Microsoft first announced DirectStorage for PC back in 2020, with Forspoken being the first game to officially support it in early 2023. However, it wasn’t until Ratchet & Clank: Rift Apart was released later that year that we saw the full DirectStorage suite in action. Ratchet & Clank: Rift Apart was the first title to ship with GDeflate compression and support for GPU decompression of assets – a task that had previously been the responsibility of the CPU.

In theory, this should have facilitated more seamless streaming of assets with smoother performance, as the feature aimed to reduce the CPU bottleneck associated with the decompression of assets during gameplay. In practice, the opposite happened, in particular on Nvidia GPUs.

What is DirectStorage?

DirectStorage on PC aims to bring many of the benefits of the fast storage technology used in the PS5 and Xbox Series X|S. Its purpose is to allow games to make full use of NVMe SSDs with minimal CPU overhead, allowing for reduced load times, faster asset streaming, and larger, more dense worlds in games. DirectStorage 1.1 also added support for GPU decompression, which would shift the burden of decompression of game assets from the CPU to the GPU. This amplifies the amount of data that can be transferred through the SSD -> RAM -> VRAM pipeline.

Unlike CPUs, GPUs have thousands of cores, and they are also very efficient at performing repeatable tasks in parallel. GDeflate is a data-parallel compression scheme that is specifically optimized for GPU decompression.

GDeflate has two levels of parallelism. First, the original data stream is split into 64 KB tiles, and each time is compressed separately. If the CPU does the decompression, then each tile can be decompressed by a different thread. If the GPU does the decompression, then each tile can be decompressed by a single thread group. Second, the data is arranged within a tile so that many lanes within a thread group can decompress that tile in parallel. GPU decompression not only saves CPU cycles, but also saves system interconnect bandwidth and on-disk footprint since the data remains compressed until it reaches VRAM.

One benefit of data moving at a faster rate through this pipeline is that, theoretically, you would need to hold less data in system memory at any point in time, which could be extremely helpful given the skyrocketing prices of DDR. This will be especially true if game developers start to lean even more into the high bandwidth of NVMe storage. Another benefit, which is especially evident in Ratchet & Clank, is that textures load in faster with DirectStorage enabled. As you can see above, with it disabled, you get blurry textures until the higher resolution textures load in.

What is the problem?

When Ratchet & Clank: Rift Apart launched, many users reported that disabling DirectStorage by removing the dstorage.dll files from the game folder (and therefore falling back to CPU decompression) resulted in better performance – particularly in terms of more stable frametime. The issue affected mainly Nvidia GPUs, including the 4090 and 3090.

In early 2025, Marvel’s Spider-Man 2 was released on PC with DirectStorage and GPU decompression support. Once again, there were reports of improved performance when disabling DirectStorage and falling back to CPU decompression on NVIDIA GPUs. When testing on a 4090, I saw increases of 18-25% in 1% lows in Spider-Man 2 when disabling DirectStorage, depending on resolution. These GPUs struggled to handle both rendering and decompression tasks simultaneously.

Do Blackwell GPUs suffer from this issue?

My initial tests were performed on a 5090, but I was pleasantly surprised to see that leaving DirectStorage enabled no longer tanked performance. However, the 5090 is an absolute beast, so this was not necessarily a sign that the Blackwell architecture is better suited to handle GPU decompression. For that, we need to test more Blackwell GPUs across the stack. Note that AMD Radeon GPUs never experienced this issue, so we will only be testing Nvidia cards in this article.

Test system

• AMD Ryzen 7 9800X3D
• 64GB (2x32GB) G.SKILL Flare X5 DDR5 @6200 MHz CL30
• Crucial T700 Gen5 SSD
• Asus ROG STRIX B850-F Gaming WiFi
• Corsair Nautilus 360 RS AIO Cooler

As you can see in both Spider-Man 2 and Ratchet & Clank, the RTX 5090 does not skip a beat with DirectStorage/GPU decompression enabled at any resolution. In fact, we now see some gains in average framerate and 1% low in some cases.

Below, we test the RTX 5070 in the same games at the same settings at 1080p and 1440p.

Similarly, the RTX 5070 sees some nice gains in 1% lows with DirectStorage enabled. At 1440p, the GPU load throughout the Spider-Man 2 benchmark is over 98%, so we are GPU-bound, and yet the 5070 has no issues rendering and decompressing assets simultaneously during gameplay.

Now for an even bigger test. Can the RTX 5060 maintain the same level of performance when it has to render and decompress assets on-the-fly?

Indeed, it can. At 1080p, the GPU load throughout the Spider-Man 2 benchmark run is over 98%, which means that even when GPU-bound, the RTX 5060 does not lose any performance when it is tasked with decompressing assets during gameplay.

By contrast, you can see below how the RTX 4060 handles GPU decompression.

The performance degradation can be quite significant in terms of the 1% lows, which is indicative of how smooth and stable the game feels when traversing the game world. In contrast with the RTX 5060, the 4060 struggles when it is tasked to decompress assets.

However, GPU decompression is working as intended on the 40-series from a texture streaming perspective. Textures load in on time, just as they do on the 50-series.

Possibly the most interesting result, however, is what we see on a 4060 laptop.

Laptop specs

• RTX 4060 Laptop
• Intel Core i7-13620H
• Gen4 SSD
• 16GB DDR5

As you can see, even on a system with a lower-end CPU, we still experience a drop in performance when shifting the decompression task from the CPU to the GPU on a previous-generation RTX card. This is true even at 720p when we are CPU-bound (the GPU load in Spider-Man 2 throughout the benchmark is 76%.)

From the tests, it is clear that Blackwell GPUs across the stack experience no performance degradation – from the 5090 all the way down to the 5060. Meanwhile, previous-generation GPUs still struggle with GPU decompression, at least the ones tested here.

Why do Blackwell GPUs handle GPU decompression better?

We only tested a limited selection of hardware and scenarios, but our results do show a few clear tendencies. It’s not entirely clear why the RTX 50-series handles GPU decompression better than their predecessors. The Blackwell data center GPUs have a dedicated decompression block, but there is nothing in the consumer Blackwell whitepaper that indicates the consumer GPUs have such a block.

One possible explanation is the addition of an improved scheduler in the Blackwell architecture. Despite its name – AI Management Processor (AMP) – the scheduler can be used to improve any asynchronous workloads running simultaneously with other graphics workloads, not just AI-related tasks. The AMP is implemented using a dedicated RISC-V processor, which isn’t anything new for NVIDIA GPUs, as RTX cards since Turing have featured RISC-V-based GPU System Processors. What does appear to be new is the fact that it was built specifically for Windows Hardware-Accelerated GPU Scheduling (HAGS), which allows the GPU to handle its own memory more efficiently without having to rely on the CPU.

According to the whitepaper, AMP matches the Microsoft architectural model that describes a configurable scheduling core on the GPU through HAGS. The AMP appears to be smarter and more efficient than previous generation schedulers, and the faster and more efficient scheduling of asynchronous workloads that it facilitates could be what we are seeing with GPU decompression on Blackwell GPUs.

In my Ryzen 7 9850X3D review, I called AMD’s latest gaming chip “a 9800X3D in a trench coat.” It was a quip at AMD that, although technically the fastest gaming processor around, the new CPU was only 3.3% faster than the Ryzen 7 9800X3D, despite selling for anywhere from $40 to $70 more. It’s a small margin that isn’t worth the extra money, but it’s still a consistent one. Say what you will about the 9850X3D, but it is technically faster than the Ryzen 7 9800X3D in games where you aren’t completely bound by the GPU.

But PBO (Precision Boost Overdrive) changes that dynamic. For our CPU reviews, we manually disable PBO to keep testing consistent. AMD’s PBO is dynamic and allows the processor to eke out a bit of extra performance when thermal and/or power conditions allow. It’s an uncontrolled variable in our reviews, voids your warranty, and is dependent on silicon and your specific setup, so we leave it off. However, it’s very easy to turn on. And the Ryzen 7 9800X3D with PBO turned on looks an awful lot like the Ryzen 7 9850X3D.

I never expected wonders out of the Ryzen 7 9850X3D. It’s identical to the Ryzen 7 9800X3D, short of a 400MHz boost in maximum clock speed. It’s possible to hit those kinds of speeds on a Ryzen 7 9800X3D, though not without a lot of manual tuning and some luck from the silicon lottery. Unless you’re an overclocking enthusiast with patience and a bit of luck, you shouldn’t expect 5.6GHz out of the Ryzen 7 9800X3D, while the Ryzen 7 9850X3D can hit those speeds out of the box.

Although you shouldn’t expect 5.6GHz from the Ryzen 7 9800X3D easily, you can still overclock it with PBO. Turn on PBO, let your motherboard determine the power limits (or turn them off), and add a 200MHz positive manual boost clock override. That’s something just about any Ryzen 7 9800X3D can do, assuming you have a decent CPU cooler. With just an extra 200MHz, the Ryzen 7 9800X3D looks nearly identical to the Ryzen 7 9850X3D in games.

We retested the Ryzen 7 9800X3D and 9850X3D with PBO turned on and a positive 200MHz boost clock override for both CPUs, to see how that impacted gaming performance, efficiency, and clock speeds. In short, the Ryzen 7 9800X3D can make up the thin margin between it and the Ryzen 7 9850X3D. AMD’s latest chip, however, has little to gain from even more clock speed, at least in games.

Before getting into the results, it’s worth reminding everyone that using PBO will void your warranty. We’re technically overclocking here, and that’s not covered by AMD’s warranty. If you’re concerned about that, just buy the Ryzen 7 9850X3D. Look at the extra $40 to $70 as the cost of an extended warranty.

Overclocking the Ryzen 7 9850X3D and 9800X3D with PBO

You can do a lot with PBO, and even more if you combine it with tweaking additional settings in your BIOS. I kept everything simple here, mainly because you don’t need to do much to get the Ryzen 7 9800X3D to perform like the Ryzen 7 9850X3D in games. In the BIOS, I turned on PBO to the “Advanced” mode — “Auto” is the default on most motherboards — and set a positive 200MHz boost clock override; the maximum allowed in PBO2. This isn’t a static 200MHz overclock. Rather, it extends the upper bound of PBO to allow up to 200MHz extra if thermal and/or power conditions allow.

I let my motherboard, in this case an MSI MPG X870E Carbon Wi-Fi, determine the power limits. That’s it. I didn’t do any manual tuning, undervolt with Curve Optimizer, or manually tune the memory beyond the 6000 MT/s EXPO profile pre-loaded on the kit.

From there, I ran Prime95 for 30 minutes to confirm the machine was stable and ran a 10-minute pass in Cinebench 2026 to check peak clocks on a single core and a single thread. That ended up being important because, as you’ll see in my tests, games don’t demand peak clock speeds from these chips — certainly not up to the 5.85GHz allowed to the Ryzen 7 9850X3D with PBO enabled.

Again, you can go a lot further than I did here. The scalar, for instance, will allow you to increase the length of the boost, though at a cost to CPU longevity. The whole point is seeing what you can quickly and easily achieve on any Ryzen 7 9800X3D. This isn’t a one-click overclock. It’s a two-click one.

I’ll list the full list of components for the test bench at the end of this piece, so skip ahead if you want to see what the bench looks like. The important note is that I tested with an RTX 5090 Founder’s Edition to remove any potential GPU bottlenecks.

With an extra 200MHz, the Ryzen 7 9800X3D gained 2.7% on average over its stock performance, leaving only 0.6% on the table compared to the stock Ryzen 7 9850X3D. Meanwhile, the Ryzen 7 9850X3D gained a mere 0.9% with an extra 200MHz from PBO. Outside of average performance, the Ryzen 7 9800X3D gained quite a bit on the 1% lows — 5.1%, specifically, while the Ryzen 7 9850X3D saw a more modest 2.7% improvement in 1% lows.

Some of the other geomeans are interesting, as well, most notably clock speed. In games, at least, you see the extra 200MHz from PBO with the Ryzen 7 9800X3D, but you don’t see it with the Ryzen 7 9850X3D. Actually, the Ryzen 7 9850X3D with PBO had a slightly lower average clock speed (though only by an inconsequential 18MHz).

Power tells a similar story, with the Ryzen 7 9800X3D picking up an extra 20W with PBO on (24% higher than stock), but the Ryzen 7 9850X3D tops out right at 106W across both the stock and PBO passes. Despite the Ryzen 7 9850X3D coming out with marginally more power draw, the Ryzen 7 9800X3D with PBO enabled actually ended up the hottest during our game testing, jumping nearly 16% compared to stock behavior.

That suggests what we all expected, which is that the Ryzen 7 9850X3D is a binned 9800X3D. We’re dealing in very tight margins here, however.

A Plague Tale: Requiem Benchmarks — Ryzen 7 9850X3D PBO

A Plague Tale: Requiem is an interesting game to look at. Although it clearly scales with clock speed, the game seems to benefit far more from the unshackled power available through PBO. The Ryzen 7 9800X3D gains 4.4% with PBO compared to stock performance, outpacing the stock Ryzen 7 9850X3D. AMD’s latest chip, however, gained an impressive 7.3%, and with much better 1% lows in tow.

Baldur’s Gate 3 Benchmarks — Ryzen 7 9850X3D PBO

Baldur’s Gate 3 sees virtually no scaling outside of the stock Ryzen 7 9800X3D. The results for the Ryzen 7 9850X3D (both stock and PBO) and the Ryzen 7 9800X3D with PBO are slightly different, but they’re functionally identical, evidenced by the consistent 1% lows. Looking at clock speed, you can see the game tops out around 5.45GHz, with the stock Ryzen 7 9850X3D actually achieving the highest clock speed at 5.5GHz. This likely has to do with all-core clocks with PBO, suggesting there might be some minor performance gains in this game if you tune the Ryzen 7 9850X3D with Curve Optimizer on a per-core basis.

Counter-Strike 2 Benchmarks — Ryzen 7 9850X3D PBO

Counter-Strike 2 is somewhat of a platonic ideal for this test. Scaling falls exactly how you’d expect it to, with both the Ryzen 7 9800X3D and 9850X3D gaining around 1.5% in average frame rate with an extra 200MHz via PBO. The real winner here for both chips is consistency. The Ryzen 7 9800X3D gained 16.6% in 1% lows with PBO, while the Ryzen 7 9850X3D gained nearly 21%.

Cyberpunk 2077 Benchmarks — Ryzen 7 9850X3D PBO

Cyberpunk 2077 is heavy on your GPU, so it’s no surprise that the stock Ryzen 7 9800X3D and 9850X3D put up almost identical performance. Looking at clock speed, you can see average clocks refuse to budge beyond about 5.4GHz, explaining the stonewall these chips are running into. Regardless, every situation outside of the stock Ryzen 7 9800X3D demands far more power and cooling potential, and for very little performance gain.

Doom: The Dark Ages Benchmarks — Ryzen 7 9850X3D PBO

Doom: The Dark Ages is heavy on the GPU, as well, and even moreso than Cyberpunk 2077 due to its always-on ray tracing. The idTech 8 engine shows good CPU scaling between different chips, but the margins are very tight here. You’re drawing more power and generating more heat, but the stock Ryzen 7 9800X3D is where you want to be for the best efficiency and largely similar performance.

F1 2024 Benchmarks — Ryzen 7 9850X3D PBO

Like Counter-Strike 2, the results here fall where you’d expect them, though with less consistent steps in between. The Ryzen 7 9800X3D saw a decent 3.2% improvement with PBO, while the Ryzen 7 9850X3D only saw a 0.5% improvement. The consistent jump was in 1% lows, with both chips improving by nearly 5%.

Far Cry 6 Benchmarks — Ryzen 7 9850X3D PBO

Far Cry 6 is more inconsistent than some games in our test suite, requiring five passes for each chip to get usable results. Here, you can see the median result for the stock Ryzen 7 9850X3D was actually a touch higher than the PBO version; that’s just what you get with this title sometimes. With less than 1% between them, we’re looking at functionally identical performance, especially approaching 300 fps. The Ryzen 7 9800X3D held up better, jumping 3.7% with PBO and overcoming the inconsistency inherent in this game.

Final Fantasy XIV Benchmarks — Ryzen 7 9850X3D PBO

Final Fantasy XIV is one of the titles where the stock Ryzen 7 9850X3D showed big performance gains over the stock Ryzen 7 9800X3D, so it’s no surprise that PBO helps here. What is surprising is how the performance tapers off. The Ryzen 7 9800X3D saw a solid 5.8% jump with PBO, and a 7.2% increase in 1% lows. The Ryzen 7 9850X3D, meanwhile, only saw a 2% improvement with 1% lows in lockstep.

Flight Simulator 2024 Benchmarks — Ryzen 7 9850X3D PBO

Flight Simulator 2024 is a more aggressive example of what Final Fantasy XIV shows. The Ryzen 7 9800X3D jumped by 5% with PBO turned on, but past that point, the scaling completely disappears. The performance is functionally identical between the Ryzen 7 9800X3D with PBO and both versions of the Ryzen 7 9850X3D.

Marvel’s Spider-Man 2 Benchmarks — Ryzen 7 9850X3D PBO

Spider-Man 2 is a game that’s notoriously heavy on CPUs, but the scaling with clock speed is minor. Even the Ryzen 7 9800X3D jumped just 1.7%, while the Ryzen 7 9850X3D saw virtually no change in performance. The big change came in power consumption, as is the case with every game I tested. Here, however, you’re just not getting much extra performance for the higher power draw.

Hitman 3 Benchmarks — Ryzen 7 9850X3D PBO

The rest of the benchmarks here tell mostly the same story. We include them to keep our test suite rounded, but we’ll let the data speak for itself.

Hogwarts Legacy Benchmarks — Ryzen 7 9850X3D PBO

Minecraft RTX Benchmarks — Ryzen 7 9850X3D PBO

Monster Hunter Wilds Benchmarks — Ryzen 7 9850X3D PBO

Starfield Benchmarks — Ryzen 7 9850X3D PBO

The Last of Us Part One Benchmarks — Ryzen 7 9850X3D PBO

Quick and Dirty Gets the Job Done

As mentioned, flicking on PBO and setting a boost override of +200MHz is about as easy as overclocking gets. Even within PBO, you have access to Curve Optimizer and Curve Shaper to squeeze as much performance out of the silicon as possible, either across all cores or on a per-core basis. And that’s before manually tuning your memory.

PBO is dynamic, so the boost override of an extra 200MHz simply allows the CPU to scale to higher frequencies if the workload demands it. It’s not a flat overclock, and that behavior is important here. As you can see consistently throughout both stock and OC performance, the Ryzen 7 9850X3D and 9800X3D top out around 5.4GHz on average in games, with only a handful of cases where they crack the 5.5GHz barrier. These chips are capable of hitting higher clocks with PBO, which I confirmed before running any games. It’s just that gaming as a workload doesn’t demand those clock speeds from these two chips.

Even lightly-threaded games like Counter-Strike 2 don’t see a linear increase as the clock speed increases. In games that scale better to higher thread counts, like Cyberpunk 2077, the differences are dulled further. And in GPU-bound titles like Doom: The Dark Ages, they disappear entirely.

Although only $20 separates the Ryzen 7 9800X3D from the 9850X3D at list prices, the 9800X3D has seen price drops in response to AMD’s latest CPU. At the time of writing, it’s available for $443, and over the past week, I’ve seen it for as little as $430. Especially below $450, it’s hard to justify the Ryzen 7 9850X3D over the 9800X3D when the latter offers almost identical performance with PBO enabled. And, although the Ryzen 7 9850X3D can climb higher, that extra clock speed doesn’t amount to much in games — in our suite, it amounts to 0.9%.

Above, you can see the test bench we used for this batch of testing, which is identical to the system we used for our Ryzen 7 9850X3D review in both hardware and software. As usual, we tested with EXPO/XMP turned on, ReBAR enabled, and Virtualization-Based Security (VBS) turned off.

Intel is reportedly preparing to launch its Arrow Lake refresh, aka 16th Gen, next month with at least three SKUs present at launch. One of those is the Core Ultra 250K Plus, the successor to the existing Core Ultra 245K, with a bumped-up core config and tuned clock speeds. While we've seen benchmarks leaks for other ARL-R chips before, the 250K Plus has just appeared for the first time on Geekbench.

This processor ended up scoring 3,113 points in the single-core test and walked away with 15,251 points in the multi-core test. Those are some pretty solid numbers and in line with what we expect from a new Core Ultra 5 desktop processor. For some reason, Geekbench's CPU database doesn't include the 245K so we don't have a direct point of comparison, but we can still search for individual runs on the platform.

Most of the 245K results listings show a single-core score under 3,000, but multi-core is almost always between 17,000-18,000, which means the Core Ultra 250K Plus tested here was ahead in one place but trailing in the other. This leaked run was conducted on an Asus Prime Z890-P WIFI motherboard with 32GB of RAM, and the processor boosted up to 5.3 GHz during testing. That lines up with the previous leak from last year.

The Core Ultra 250K Plus is supposed to be an 18-core processor with 6 P-cores and 12 E-cores (up from 6P+8E on 245K). It features a 100 MHz increase in both of those core boost clocks, along with 100 MHz decrease in the E-core base clock speeds. Like other Arrow Lake refresh chips, it has support for native DDR5-7200 and will use the existing LGA 1851 socket, yet there are new motherboards planned.

This will be Intel's last hurrah in the desktop segment until Nova Lake is expected to debut at the end of this year, marking a true next-gen leap similar to what Panther Lake has done on mobile. Recently, prices and embargo dates for these ARL-R chips leaked out and showed a very enticing picture, while confirming that the Core Ultra 9 SKU is dead for this generation.

Minisforum let me go hands-on with its CES 2026 Innovation Award-winning Mini Workstation, so to get a grasp of what the diminutive new Minisforum MS-02 Ultra could offer, I decided to see if it could replace my aging but capable desktop PC.

This is a true David vs Goliath battle, at least in terms of physical presence. The sub-5-liter MS-02 Ultra battles with my much larger desktop system packed into a 55-liter Fractal Define S PC case. However, the result of the clash wasn’t as clear-cut as I expected, making me wonder what kind of PC I should be using daily in 2026.

Before we get into a specifications table, where you can see exactly how these two systems compare feature-by-feature, it is important to highlight the physical differences. The MS-02 Ultra is an incredible 11 times smaller than my Define S-desktop system. Obviously, such a small size means there are sacrifices to be made in terms of expandability, configurability, thermals, and noise. But those limitations are counterweighted by the benefits of the device being much physically smaller. Dare I say, the Minisforum is kind of portable, and even cute, which is not something I could say about the Define S PC.

The table below makes clear the specs of the Minisforum MS-02 Ultra workstation against those of the system it will have to battle on my desk.

CPU and GPU showdown

The Intel Core Ultra 9 285HX should be much more capable than the i5-14400F that I bought at Amazon for $115 last year. It tempts with the advantages of its modern Intel architecture, higher clock boost, and easily outguns my Raptor Lake Refresh chip on core count. Moreover, the new platform brings quite a lot of modern niceties. Particularly welcome, from my perspective, are the trio of really fast USB4 (including 2x USB4 v2) ports, all with DisplayPort capabilities.

A distinct advantage I expected in moving to the MS-02 Ultra was from the increased core count using Intel’s more modern architecture. Indeed, CPU-heavy benchmarks would show the new chip could convincingly eclipse the old mid-range Core i5-14400F in both single- and multicore workloads, given sufficient power supply and cooling capacity. But, we'll see how that translates to ‘Mark’s world’ of computing in the benchmarks section, below.

I’ll stress again that a brief check of the specs, ahead of any actual comparison testing, may suggest the little Minisforum would win hands-down in CPU-focused processing tasks, but the best consumer-grade GPU I could get to fit inside its sub-5-liter chassis would be crushed by the triple-fan behemoth in my ATX desktop. Indeed, that is sort of what happened. There are nuances, though, especially as I don’t really play any AAA or eSports PC games that launched in the 2020s, or dabble in local AI.

The MS-02 Ultra runs using the integrated graphics on the Core Ultra 9 285HX by. default. However, there’s a PCIe5.0 x16 slot, and incredibly, there’s room to install a two-slot low profile (LP) graphics card. This, plus the spare 8-pin connector from the built-in 350W PSU, allowed me to install a Gigabyte RTX 5060 OC LP GPU.

Nvidia’s RTX 5060 seemed to be the best graphics card choice for me, as a generalist/hobbyist PC user, though I’ve seen other early adopters of the MS-02 Ultra install cards like the Intel Arc Pro B50 (16GB), Nvidia RTX Pro 4000 Blackwell SFF (24GB), and the RTX 4000 SFF Ada (20GB). These offer far more VRAM but can cost a lot more, too. But thanks to this mini workstation allowing for GPU upgrades, perhaps the RTX 6060 or RTX 7060 will arrive one day with more than 8GB of VRAM on board, and in a low-profile form factor. Or maybe AMD will make a performant LP form factor card in the meantime.

More on specs

I chose my Fractal Define S-housed desktop's motherboard as it was one of the few B760 chipset boards with a USB-C 20 Gbps port as standard at the price point. So, getting three times more USB-C ports, with two of them supporting up to 80 Gbps as standard, on the compact MS-02 Ultra was a welcome platform boost. I’m sure there is potential here for some very useful docking or even eGPU use-cases.

I have a pair of Crucial X10 Pro 4TB devices for keeping games, TV, and movie libraries portable. My testing confirmed the USB4 v2 ports on the front work at full USB 3.2 Gen 2×2 (20 Gbps) speeds with the X10 Pro. Some USB4 controllers surprisingly don’t support Gen 2x2 and would fall back to 10 Gbps mode. My X10 Pros run about a third of the speed of the built-in M.2 SSD provided by Minisforum, but I feel they are good for external, highly portable units.

An extra two 25 GbE network ports and two of the available M.2 SSD slots are provided by a pre-installed low-profile card in the MS-02 Ultra’s PCIe 4.0 x16 slot. However, I chose to replace that with a cheap USB Type-A 3.0 card with four ports, as I really like to have lots of USB ports. My old desktop used a USB adaptor for Wi-Fi 6 and didn’t have Bluetooth. The MS-02 has built-in Intel BE200 Wi-Fi 7 and BT 5.4 card.

There are three PCIe slots in the MS-02 Ultra, but adding the GPU takes up the room where two single-slot add-in-cards would otherwise fit. That compares poorly with my Define S desktop, which has five slots. With a bulky GPU fitted in that ATX system, it was still possible to add another three single-slot PCIe cards. Again, I have one installed in there, bristling with extra USB ports.

Preparing for my personal benchmark battle

The most demanding things thrown at my PC are video editing and games like The Division 2 and Far Cry 5. The Cyberpunk 2077 benchmark is just there for an added reference point, (I bought it, but I’ve not actually got around to playing it yet.) The same goes for Shadow of the Tomb Raider. I also like playing in Borderlands GOTY Enhanced Edition, Command & Conquer Generals, and Amiga Forever – but those are so old and lightweight they weren’t worth testing for this performance comparison.

Most of the time, my computer will be used for little more than multiscreen word processing and browsing, communicating with work colleagues, and a touch of ancient Photoshop and Illustrator dabbling. These help me prepare artwork for here on Tom’s Hardware, some YouTube stuff, as well as laser cutting and engraving. Also, Photoshop has recently been replaced by the Affinity suite.

So, without further ado, let’s get into those benchmarks, which unashamedly revolve around my personal general-purpose computer use, with a few gaming classics thrown in. I’ve also included a ‘tweener’ system, where my large desktop was fitted with the incongruously tiny RTX 5060 LP card.

Minisforum MS-02 Ultra productivity performance

The comparison table above doesn’t offer much in the way of surprises. The MS-02 Ultra with Intel’s Core Ultra 9 285HX does exceedingly well in CPU-intensive tests, particularly those that take advantage of the abundance of available threads. We see the 285HX is a star performer in Cinebench R23, with single-thread scores similar to desktop processors like the Core i9-14900K and Ryzen 9 9950X. For multicore, you are also good to compare with the Ryzen 9 7950X3D and i9‑14900F, but the unlocked i9‑14900K starts to pull ahead convincingly.

I also think the 7-Zip (de)compression benchmark shows the MS-02 mini workstation in a good light. Handbrake 4K transcoding wasn’t much better compared with my old 14400F, though. Meanwhile, PassMark Performance Test, a mix of CPU, graphics, memory, and disk tests, showed some worthwhile performance benefits from the Minisforum – when the RTX 5060 was installed.

Minisforum MS-02 Ultra gaming performance

It is almost moot to highlight that the Define S desktop with Radeon RX 9070 would dominate in 3D games and 3D benchmarks. Nevertheless, there were some interesting results showing that the Ultra 9 285HX / RTX 5060 didn’t always beat the i5-14400F / RTX 5060 due to titles simply being ‘GPU-limited.’ An interesting outlier was the crazy-town USA-set Far Cry 5, where Ubisoft’s Dunia Engine 2 (a modified CryEngine) seemed to prefer the desktop Raptor Lake Refresh under the hood. All the games were tested in 1440p, with no upscaling or frame-gen tech applied.

Minisforum MS-02 Ultra power consumption and noise

Now we move into an area of both strength and weakness for the Minisforum MS-02 Ultra. In my benchmarks table, above, I also included power and noise readings. Starting with the good news, the MS-02 requires less power while idling and doing light tasks – like writing this. It also uses about 120W less when gaming, and all the titles I enjoy play smoothly enough on the RTX 5060 system at high to ultra settings at 1440p without the use of any scaling or frame gen tech.

Some bad news for the mini workstation is that it is more audible when idling and under load, in my experience. Initially, I measured both systems from about arm’s length distance. My Define S desktop uses large Noctua fans and sits under the desk, and in the comparison results, you see that even using the tiny RTX 5060 didn’t increase the noticeable system noise – it is hard to be 100% confident about the 1dBA difference observed.

Having the MS-02 Ultra at arm's length from my seated position, but on the desk top, the noise level was quite high at 48 dBA under load, as per the above chart. However, sitting it under the desk, where the big DIY desktop PC was previously located, reduced the max noise level to 41 dBA. But it's a far smaller system, and when positioned in the space vacated by the Define S, the Minisforum front panel moves beyond arm's reach under the desk. It is necessary to bend quite low from the chair to reach the power button and front I/O, which isn't convenient. That's why I decided to keep it on the desk.

I also note that the Minisforum BIOS has various fan modes (quiet, balanced, and performance) and manual tweaking of the trio of system fans. I ran it at ‘Balanced,’ so there’s probably some tuning that could be done for better noise performance.

Bottom line

In some many ways, the Minisforum MS-02 Ultra held its own against my old home-built Goliath.. The pint-sized challenger does everything I need, and some more. There are clear advantages to be had in CPU-heavy workloads. In GPU-heavy tasks and gaming, there’s clearly enough juice for my particular needs today. If you think about the future, then unlike many other mini PCs, this one can get upgraded with a faster next-generation low-profile graphics card (and other compact PCIe cards) if and when they arrive at an affordable price.

I’ll continue to work on configuring the fans, with an eye on processor and memory temperatures, as the louder fans are the one aspect of the MS-02 Ultra that pokes me in the ribs. It may also be repositioned further back on the desk or under it to scale back its audible presence. The fan noise isn’t high-pitched, and I sometimes forget it, but less is more.

We must also talk about the price of the Minisforum MS-02 Ultra mini workstation. As of the this writing, the Intel Core Ultra 9 285HX model with 2x 25 GbE PCIe SFP+ NIC, 32GB DDR5, plus a 1TB M.2 PCIe4x4 SSD, and Windows 11 Pro is priced at $1,583.90 from Amazon.com. You can also go barebone (no RAM, SSD, or OS) for $1,229.90 using a $300 discount coupon on the page, which could be a great option if you have a spare DDR5 SO-DIMM, M.2 SSD, and can source your own OS. As with all PCs at this time, prices are very volatile.

The Gigabyte GeForce RTX 5060 OC LP 8G I bought to go in this system's PCIe ​5.0 x16 slot is currently $359 at Newegg. This was a necessary extra to indulge in the type and quality of gaming fun that I am used to.

Thus, for the Minisforum MS-02 Ultra 32GB/1TB with RTX 5060 LP graphics card installed, as tested, today's total price would be $1,942.90. Things have gotten a little pricier since I began my tests in early January, but that's the way the PC market is going.

We could talk about the comparison between the MS-02 Ultra build/cost and a desktop, with the latter offering much better value and expandability or upgradeability. However, getting something this performant yet compact, with this level of I/O spec and expandability, is going to be impossible with off-the-shelf parts. In this way, and especially for those with a use for the 2x 25 GbE PCIe SFP+ NIC, this mini workstation is a uniquely attractive package.

It's a happy day for game benchmarking enthusiasts, including our very own. After languishing for a year since its previous update, Nvidia's FrameView 1.7 is now available for download, bringing with it a host of fresh improvements.

Perhaps in keeping with Nvidia's recent focus on boosting frame rates via frame generation, FrameView 1.7 claims to offer "improved accuracy with FPS calculation in situations where games have very high frame rates (800+ FPS)." That improved accuracy at is likely only relevant for titles where MFG is enabled at 6x and lower resolutions, or with older games that can run at ultra-high speeds on modern hardware.

FrameView 1.7 also includes the option to better customize the in-game overlay that runs alongside the app. Users can now adjust the metrics shown in the overlay from just FPS to the full suite of FPS, 1% low frame rates, PC latency, and GPU and CPU clocks, an improvement that video content creators and more casual users might appreciate for reducing visual clutter.

Additional improvements include eliminating a memory leak that led to crashes in long benchmarking sessions with Reflex-compatible games, and fixes to a rare instance where using V-sync, G-Sync, and DLSS Frame Generation could lead to spurious results. Version 1.7 should also properly save the CSV data files if the game closes or crashes when FrameView is running.

Nvidia also corrected some game-level compatibility issues across The Finals, Arc Raiders, Starfield, Black Myth Wukong, and Battlefield 6 that should make FrameView more stable and useful with those titles.

Although software like FrameView doesn't necessarily need to be updated every other week, FrameView 1.6 (released in February 2025) was getting long in the tooth , and it's arguably been multiple years since the utility had a rich feature update.

Judging by forum posts across the internet, many testers and enthusiasts have since moved on to alternatives like CapFrameX thanks to its instant data visualization from log files and baked-in support for the RTSS overlay. Our own Jeff Kampman continues to use FrameView due to its per-frame power reporting integration with Nvidia's PCAT hardware power-monitoring tool, but many other members of the media have since moved on to greener pastures.

Also, given there's no indication otherwise, FrameView 1.7 likely still relies on PresentMon 2.2 as its foundation, a version that is now 1.5 years old, too. The most up-to-date PresentMon 2.4.1 includes many bug fixes and adds many additional metrics, including a number of telemetry improvements for Intel Arc cards. Assuming it hasn't already, Nvidia will likely roll a newer version of PresentMon into a future version of FrameView, whenever it might arrive.

As a largely mature application with a small user base, FrameView probably won't see too many major changes to its basic functionality going forward, but there are some things that could still be added. Its PC Latency reporting is the most solid software estimation of that figure across vendors, in our experience, and that data is extremely handy when making performance tuning decisions with MFG.

If Nvidia made it easier to split out generated frames from native frames, as only the Steam performance overlay reliably does right now, FrameView would likely become more interesting as a general-purpose performance monitoring tool in the framegen era.

If you're looking to try out the improvements yourself, hit the FrameView download page, and perhaps check out the updated user manual to learn about all the new goodies.

Windows 11 gets a bad rep in the community because of its higher baseline overhead, stringent hardware requirements, UI regressions, and more - not to mention the forced Microsoft hooks that keep getting worse by the day. Moreover, when placed in a rather unscientific test by TrigrZolt, comparing six different generations of Windows with each other, it placed dead last in pretty much every individual test, though the situation is a bit more nuanced.

Six Lenovo ThinkPad X220 laptops were used in the test, featuring a Core i5-2520M CPU and 8GB of RAM, with a 256GB hard drive — running the latest versions of Windows XP, Windows Vista, Windows 7, Windows 8.1, Windows 10, and Windows 11. That setup alone should tell you how the methodology employed here is skewed toward favoring older software. Windows 11 isn't even officially supported on these components.

Regardless, the experiment begins with a startup test, where Windows 8.1 booted up the quickest, while Windows 11 was the slowest. Both of these versions (and Windows 10) have Fast Boot capability that older Windows editions lack. In the video, we also see that Windows 11 struggles to load the taskbar for a bit, an infamous quirk of the OS that's been heavily scrutinized ever since launch.

Then comes the storage test, where it's Windows XP that takes the cake, with only 18.9GB of space taken up for all the apps installed. The same number of programs, along with Windows itself, took 37.3GB of hard drive real estate on Windows 11, so there's definitely a lot of extras there. But Windows 11 actually came third here, behind Windows Vista, at 37.8GB, and the revered Windows 7, at a whopping 44.6GB.

Next up is RAM management where Windows XP is the winner once again, consuming only 0.8GB of system memory at idle, while Windows 11's appetite grew to 3.3GB on average; it jumped to 3.7GB at one point. This is because of the added resources the OS loads in the background, including persistent telemetry.

Older hardware with less RAM, therefore, will be more susceptible to sluggishness on Windows 11. Keep in mind, TrigrZolt is also running a system with a hard drive, which are outdated at this point regardless of your operating system loyalties. Any modern system with a decent CPU and NVMe SSD will likely mask over the general inefficiency Windows 11 shows, plus options like debloat tools and Xbox FSE can further help here.

Now we move on to the second part of the RAM management test, where the YouTuber loaded as many browser tabs as possible before the memory hits 5GB of utilization. Since Firefox and Chrome don't load webpages properly anymore on archaic Windows versions, a more widely-compatible browser called Supermium was used across all devices.

Once again, Windows 11 places dead last here, only being able to load a measly 49 tabs. Compare that to the insane 252 tabs Windows 8.1 was able to load. Even the older Windows XP managed 50 tabs, and that's because it kept crashing past that number because of its paging file failing to keep up, not because it had hit the 5GB memory ceiling.

Our fourth test is for battery life and, of course, Windows 11 died first here, while Windows XP walked away with the best endurance. Though, the delta between all the devices was only about two minutes so it won't make a difference in real-world usage. All the laptops had 100% battery health, too, and the same program was run to drain them as quickly as possible.

Moving on, exporting an audio file in Audacity once again put Windows 11 at fifth place, only ahead of Windows Vista which was experiencing an unusual delay, otherwise all laptops finished around the same time. The same fate follows Windows 11 when it came to rendering a video, finishing in last, with Windows 10 taking first place. Here, Windows XP and Vista couldn't load the OpenShot Video Editor that was used, so they were disqualified.

In application opening times, Windows 11 got last place across all five programs that were tested: File Explorer, MS Paint, Calculator, Adobe Reader, and VLC Media Player. Older versions of Adobe Reader and VLC were used to ensure compatibility with all six operating systems, so there's a bit of performance left on the table, but still the native apps didn't win any awards either.

After so many consecutive losses, Windows 11 actually secured third place in one half of the web browsing test where it had to load an image, but fell to last place again when visiting the Google Images and Microsoft Account login websites. When transferring files, though, Windows 11 snatched second place, only behind Windows 10, while placing fourth in the malware scan test using MalwareBytes — Windows 7 won this one.

Finally, we arrive at our last test, which are benchmarks. Windows XP took the crown in CPU-Z's single threaded load while Windows 7 topped the multithreaded charts, and Windows 11 was fourth in both. Geekbench was also tested, but the versions are different for Windows 10/11 and the rest. Between those two, Windows 11 scored higher in single-core but lower in multicore. Among the older operating systems, Windows Vista walked away with the highest score.

TrigrZolt also tested CrystalDiskMark and Windows 11 netted a decent third place here, tied with Windows 10, while Windows XP won. In Cinebench R10 single-core, Windows 8.1 got its second victory of the day, pushing Windows 11 down to fourth place. Multicore was even worse for Windows 11 because it only beat Windows 10 by a few points to save itself from scoring last; Windows Vista walked away as the fastest.

All in all, this was a pretty devastating showing for Windows 11 where it couldn't even win a single test. However, the hardware is just so outdated at this point that it doesn't represent the Windows 11 experience faithfully. The laptops were never designed to run a modern operating system, neither does Microsoft's latest OS support this hardware.

If you were to use modern laptops that came out in the Windows 10/11 era, the results would likely be skewed the other way. The omission of an SSD, especially, is strange since that's a component every edition of Windows will benefit from, and something that truly does hold back Windows 11 here. Though, it's still a little embarrassing, considering Windows 10 didn't suffer nearly as much without it.

The YouTuber himself mentions that this test represents historic value more than practicality. He ended up giving the overall win to Windows 8.1, saying how fluid and fast it felt, despite being ridiculed at the time of its release. It's also more visually consistent with Windows 10/11 than it is with Windows 7/Vista, so it doesn't look outdated.

A better methodology would've been to use flagship — or even midrange, for that matter — laptops from every generation: custodians of that era of Windows, so that each version had the best shot at performing at its full potential. As it stands right now, while it's quite funny to see Microsoft's increasingly AI-riddled OS loose against legacy offerings, the test just wasn't set up fairly to begin with.

Heads up, tech enthusiasts everywhere, for a new version of one of your favorite benchmarks just dropped. Maxon just announced the minty fresh new Cinebench 2026, ready to beat the heck out of your processor and GPU with an updated, more demanding engine.

Cinebench 2026 uses the latest version of the Redshift engine that Maxon says ought to be six times harsher on multi-threaded tests compared to the the previous version. Predictably, the updated engine integrated into the benchmark should also better reflect the real-world demands of 3D artists everywhere, as well as offer better parity between test results and actual render tasks. Maxon notes that scores with the new engine cannot be compared with previous versions of Cinebench.

The new SMT core test should be of particular interest to testers everywhere, including those at our abode. This test specifically runs a single core test procedure on an SMT capable core, so users can get a feel for how good (or bad) their chip's SMT implementation is when compared to single thread execution.

The list of supported hardware got longer, too. The latest generations of AMD and Nvidia consumer cards are all present in the list, which also includes includes Nvidia's Ampere Altra, Hopper, and Blackwell datacenter GPUs, as well as AMD's Radeon Pro series. Owners of Intel graphics chips continue to be out of luck, as those are still not supported by the Redshift engine.

Over at the Apple orchard, the new M4 and M5 chips are now on the list, as you'd expect. While on the topic of Arm-based chips, as with Cinebench 2024, there's an native Windows-on-Arm version, along with support for Nvidia's Ampere Altra CPUs.

The new niceties do come at a price, though an expected one. For GPU testing, only cards with 8GB of more of VRAM need apply. The benchmark now supports Apple's M4 and M5 chips, and given that the M-series machines all use unified memory, Maxon says you'll need 16GB for running the GPU tests. If you're only doing CPU tests on Apple machines, you might get by with 8GB or 12GB, though the company notes that performance might not be up to spec.

Interested techies can grab Cinebench 2026 from Maxon's site and start torturing their machines. Supported operating systems include Windows 10, Windows 11, and macOS 14.7 or above.

Veteran Windows dev Dave W Plummer has run a simple performance test on all his computer lab systems and drawn up a fascinating comparison chart. Across his 25 computers, released between 1976 and 2023, he observed a huge 200,000X CPU performance delta. Moreover, it is admitted that the difference could have been much greater if the ancient Dhrystone benchmark performance code wasn’t resolutely single-threaded. The benchmark code also uses simple math, so the gains could be even more pronounced if it tested vectorized code.

Click 'See more' to unfurl the chart

If you’ve owned several generations of computers over the years, it can be interesting to step back and consider whether the newest is really that much faster than the oldest computer you owned. Veteran Windows dev Dave Plummer may have gone through a similar thought process recently, to drive him to create a code-portable Dhrystone benchmark and complete the tests. However, Plummer is blessed with an extensive computer lab featuring at least 25 computer systems spanning the 1976 vintage DEC PDP-11/34 to the Apple Mac Pro with M2 Ultra processor, released in 2023.

If you aren’t familiar with the DEC PDP-11/34, or the Apple machine, Plummer’s chart also features some of the most iconic systems / architectures in computer history. For example, the seasoned dev’s second-slowest machine is the Amiga 500 (score 1,000).

We also note the great progress from the first i486 machines through several generations of Pentium CPUs in Plummer's benchmark. These figures show Intel’s solid progress from Dhrystones in the 30,000 to 2,500,000 score regions in roughly a decade, which included with the turbulent early Windows 9X era.

Other notable results / comparisons we would pick out are the scale of the uplift delivered in Apple’s transition from Motorola 680X0 chips to Power PC. Then there is the Raspberry Pi 4B standing proudly in the chart at nearly 10,000,000 points, and approx 4x faster than an iso-clocked P4.

Plummer also seems to have a big gap in his system collection after the P4 era. Where are all the lovely Athlons, Core-Duos, and Core iX series chips? However, atop of the seasoned dev’s chart we see two modern Goliaths, powerful examples of some of the best CPUs available in recent years: the Ryzen Threadripper PRO 7995WX with 96C/192T AMD Zen4 cores, and the aforementioned Mac Pro M2 Ultra with 24-cores of Apple Silicon.

Dhrystones whystones?

In his Tweet thread, Plummer made it clear that all the charted systems ran a Dhrystone benchmark, that he wrote himself, naturally. It measures “Just raw compute on a single thread... not even taking into account multiple cores,” he explained.

If you are interested in this archaic CPU integer performance test, Plummer has made it available on GitHub. You have to grab the code and compile it for your platform. It is designed to be “specifically optimized for 2.9BSD on PDP-11 systems, while remaining compilable on modern systems for comparison.”

Running on systems newer than the 486 era, this benchmark code may fit entirely in a CPU’s cache, partly accounting for the significant uplifts seen in Plummer’s chart during this era change. Conversely, this code doesn't use the advanced types of vectorized instructions that impart huge performance uplifts on modern systems, like AVX-512, so the overall benchmark gains could be even larger in some types of tests. That said, because Dhrystone fits entirely in cache, it also doesn't test some of the primary bottlenecks in modern systems that limit performance, such as memory bandwidth or the performance of higher cache tiers.

Interestingly, as one X commenter notes, while the Amiga 500 is near the bottom of the chart, this 80s home computer with a humble 7.16 MHz 68000 CPU can still boot and open a perfectly functional word processor application faster than most modern systems today. Shove that in your Dhrystones chart…

GPU overclocker TrashBench set a world overclocking record on an Intel Arc B580 GPU with a relatively simple setup. Instead of employing exotic solutions like liquid nitrogen, used by other overclockers to hit world records, he instead slapped on a pond pump and filled the custom loop with a 50/50 glycol mix of vehicle antifreeze that had been pre-chilled in a freezer. To his surprise, the GPU hit an incredible -17 degrees Celsius (that’s 1.4 degrees Fahrenheit), allowing him to push the performance of Intel’s top-of-the-line consumer GPU. According to the YouTube video, the car-coolant-chilled graphics card achieved a 3DMark Time Spy score of 16,631 — some 12% higher than a stock GPU, breaking the world record for the B580.

The mods TrashBench made to the GPU were rather simple, with the most complicated part being 3D-printing a custom mount to secure the water block. Trashbench notes that antifreeze remains liquid until roughly -25C, and he chilled the liquid in his freezer, which took it down to -17C. From there, he ran a couple of flexible tubes into an open cooler containing the pre-chilled glycol mix and turned on the pump to circulate it, allowing the graphics card to reach sub-zero temperatures.

TrashBench broke a world record in 3DMark Time Spy when the coolant was still cold.

He also tested the Intel Arc B580 in its stock configuration, with the card hitting 2,850 MHz on air cooling alone. This delivered an average of 54, 158, and 107 FPS for Monster Hunter Wilds, Forza Horizon 5, and Cyberpunk 2077, respectively. But with the GPU running on sub-zero cooling, it achieved 3,316 MHz — some 446 MHz over stock performance or a 16.4% higher clock speed. Unfortunately, because the coolant eventually warmed up during testing, TrashBench wasn’t able to maximize its performance, but even so, the card still hit 69, 174, and 120 FPS in the three titles, giving it an average of 16% more FPS.

TrashBench's results show that with just air OC, the B580 achieved 3200 MHz, with the card hitting 60 FPS on Monster Hunter Wilds, 119 FPS on Cyberpunk 2077, and 172 FPS on Forza Horizon 5 at 1080p High settings. Although extreme cooling still seems to offer more FPS, it wasn't a huge gain over what you get without any of the hardware modifications.

Nvidia’s midrange GPUs have long been a sweet spot for gamers seeking solid performance without breaking the bank, so it's high time we revisit them. ComputerBase.de’s recent testing of the RTX 2070, 3070, 4070, and 5070 across 1080p and 4K resolutions provides valuable insights into how each generation has evolved—not just in raw performance, but in efficiency, thermal behavior, and the impact of architectural advancements. You might not be expecting any surprises in this roundup, and this might be lengthier than the average news piece, so let's take a look.

Compared at 1080p

At 1080p, the generational improvements are usually the least evident because the games are more CPU-bound. Still, the scaling is clear when you look across a wider spread of titles. In Cyberpunk 2077 with ray tracing on medium and DLSS Balanced, the RTX 2070 averages 12.9 FPS, the 3070 nearly doubles that at 23.8 FPS, the 4070 reaches 48.2 FPS, and the 5070 peaks at 61.3 FPS. That kind of progression carries into other ray-traced workloads: Oblivion Remastered climbs from 24.9 FPS on the 2070 to 63.3 FPS on the 5070, while Quake 2 RTX rises from 52.6 FPS to 166.5 FPS over the same generational span.

In more forgiving rasterized titles like Diablo II: Resurrected and Overwatch 2, which start out comfortably above 100 FPS on the 2070, the latest cards push into the 300–380 FPS range at 1080p — far beyond the limits of most monitors, though the manufacturers are always pushing for more. The takeaway is that while not every title stresses modern GPUs equally at lower resolutions, the newer architectures consistently scale better in demanding scenarios, especially where ray tracing is involved.

4k is a different story

The 4K benchmarks reveal a different story because this is where the GPUs really stretch their legs and the performance hierarchy comes into sharper focus. In Cyberpunk 2077 at ultra settings with ray tracing and FSR Balanced, the 4070 delivers 27.8 FPS, while the 5070 manages 32.0 FPS — a modest uplift, but still a generational gain. Ratchet & Clank: Rift Apart tells a similar tale, with the 4070 at 92.6 FPS and the 5070 at 119.2 FPS, more than doubling the 2070’s 19.2 FPS baseline. Horizon: Forbidden West also benefits heavily, where performance climbs from 31.6 FPS on the 2070 to 91.1 FPS on the 5070, showing how newer GPUs increasingly make 4K playable even in modern open-world titles.

The esports angle scales as well: Overwatch 2 runs at 99.3 FPS on the 2070 but hits 275.2 FPS on the 5070, proving that high-refresh 4K competitive gaming is no longer out of reach. The generational contrast is sharpest here, not only because raw throughput matters more at 4K, but also because features like DLSS—which just got updated to a much faster Transformer model—gain greater leverage.

Power and efficiency further tilt the scales because, despite being slower in absolute terms, the 4070 consistently runs with 20–25% lower power draw than the 5070, leading to 15–25% stronger performance-per-watt in titles like Cyberpunk 2077, Horizon, and Doom Eternal. In some cases, like Doom Eternal with ray tracing, the 4070 more than doubles the 2070’s performance while consuming almost the same wattage, signifying the architectural leaps. The RTX 4070 and 5070 are pretty much neck-in-neck efficiency-wise with both trading blows, but the 4070 often leads.

This small advantage actually positions it the most balanced upgrade in the lineup, particularly for players who care about thermals and acoustics as much as raw frame rates. Owners of the 2070 will find transformative gains by moving to either the 4070 or 5070, while 3070 users benefit most from the 4070’s efficiency and DLSS enhancements. But for 4070 owners, the leap to a 5070 often feels underwhelming — the higher power and thermal cost buys modest raw FPS improvements, but comes with an efficiency penalty that undermines the real-world value.

RTX 4070 is a standout

Thermal Grizzly has officially launched the der8enchtable, a professional-grade test bench developed in collaboration with overclocking legend Roman "der8auer" Hartung and Elmor Labs. First unveiled at Computex 2025, where it picked up a prestigious design award, this unique benchtable is built around an active PCB, offering functionality far beyond a simple hardware stand.

Originally designed for internal use, the project evolved into a full commercial product aimed at overclockers, content creators, reviewers, and hardware testers who need to swap components quickly without the hassle of a full case. Now, it's finally hitting the market.

The standout feature is its integrated PCB, which consolidates essential connections for storage, cooling, and peripherals directly into the bench. Cooling control is handled by three fan zones, each with 3x 4-pin headers, along with a dedicated pump zone that offers 2x 4-pin headers. Each header delivers up to 3A of power, is individually fused, and allows independent speed switching, with the option to hand over control to the motherboard.

For storage, the benchtable includes 2x onboard SATA ports and 2x additional SATA outputs, plus mounting points for 2x 2.5-inch SSDs and 4x microSD card slots for quick media swaps. Connectivity is further expanded with 4x USB 2.0 Type-A ports along with 2x USB 2.0 Type-C ports. Power is supplied through a single PCIe 6-pin connector, while a 9-pin USB header links the bench to the motherboard.

The der8enchtable comes with pre-installed ATX standoffs that can be adjusted for mATX and Mini-ITX boards, a modular frame capable of mounting a PSU or operating in compact mode, Velcro straps for cable management, and mounting points for fans or liquid cooling radiators. The underside also features RGB illumination, which can be controlled via a 3-pin aRGB header or an external controller.

By consolidating connections, the bench eliminates the need to wire multiple devices directly to the motherboard, allowing users to hot-swap drives, adjust cooling, and hook up peripherals with minimal downtime. This makes it particularly well-suited for situations where motherboards are swapped frequently — something reviewers and overclockers will immediately appreciate.

The AI revolution is upon us, but unlike past shifts in computing, most of us interact with the most advanced versions of AI models in the cloud. Leading services like ChatGPT, Claude, and Gemini all remain cloud-based. For reasons including privacy, research, and control, however, locally run AI models are still of interest, and it's important to be able to reliably and neutrally measure the AI performance of client systems with GPUs and NPUs on board.

Client AI remains a highly fluid space as hardware and software vendors work to define the types of workloads best suited for local execution and the best compute resources upon which to execute them. To help navigate this rapidly changing environment, the MLCommons consortium and its MLPerf Client working group maintain a client benchmark that's developed in collaboration with major hardware and software vendors.

MLPerf Client 1.0 has just been released with some big improvements over the prior version 0.6 benchmark. This new tool includes more AI models, supports hardware acceleration on more devices from more vendors, and tests a broader range of possible user interactions with large language models. It also sports a user-friendly graphical interface that should broaden the appeal of this test for more casual users.

MLPerf Client 1.0 can now test performance with Meta's Llama 2 7B Chat and Llama 3.1 8B Instruct models, as well as Microsoft's Phi 3.5 Mini Instruct. It also offers support for the experimental Phi 4 Reasoning 14B model as a possible example of performance with a next-gen language model with a larger set of parameters and greater capabilities than before.

MLPerf Client 1.0 also explores performance across a greater range of prompt types. It now examines performance for code analysis, as developers might commonly ask for nowadays. It can also measure content summarization performance with large context windows of 4000 or 8000 tokens as an experimental feature.

This range of models and context sizes gives hardware testers like yours truly a more broadly scalable set of workloads across more devices. For example, some of the experimental workloads in this release require a GPU with 16GB of VRAM to run, so we can stress higher-end hardware, not just integrated graphics and NPUs.

The hardware and software stacks of client AI are fluid, and the various ways that one can accelerate AI workloads locally are frankly dizzying. MLPerf Client 1.0 covers more of these acceleration paths across more hardware than before, most notably for Qualcomm and Apple devices. Here's a list of the supported paths:

• AMD NPU and GPU hybrid support via ONNX Runtime GenAI and the Ryzen AI SDK
• AMD, Intel, and NVIDIA GPU support via ONNX Runtime GenAI-DirectML
• Intel NPU and GPU support via OpenVINO
• Qualcomm Technologies NPU and CPU hybrid support via Qualcomm Genie and the QAIRT SDK
• Apple Mac GPU support via MLX

Version 1.0 of the benchmark also supports these experimental hardware execution paths:

• Intel NPU and GPU support via Microsoft Windows ML and the OpenVINO execution provider
• NVIDIA GPU support via Llama.cpp-CUDA
• Apple Mac GPU support via Llama.cpp-Metal

Last but certainly not least, MLPerf Client 1.0 now has an available graphical user interface that lets users understand the full range of benchmarks they can run on their hardware and easily choose among them.

Graphics card enthusiast TrashBench has created a video pitching the venerable flagship of the Pascal era against the ‘turnip’ of the Blackwell era, with unexpected results. The ‘Voltage first. Questions later.’ Aussie explains that he set out to “overclock a GTX 1080 Ti hard enough to embarrass Nvidia’s new RTX 5050.” However, things didn’t turn out as planned, with the 5050 grasping an unexpected triumph both at stock and with its impressive overclockability.

TrashBench’s plans regarding the GTX 1080 Ti didn’t get off to the best start with two samples of this old flagship having to be cast aside before finding one that could muster anything more than stock performance. Apparently, it took “days” to get the selected card to perform, with fine-tuning of curves, offsets, drivers, and APIs used – and a custom coolant loop – to rouse this respected champ of old.

We all get old, but TrashBench says that there was a brick-wall limit at 2.2 GHz for the best 1080 Ti he had on hand. Sadly, there are only so many rocket pods you can fit to an old silicon Zimmer frame before it topples over. At this stage of the video, it is admitted that the original “1080 Ti beats 5050 idea was dead.” Actually, things would be turned upside down.

Plucky little RTX 5050 overclocked to 3.3 GHz

Attempting to give the RTX 5050 a fighting chance, TrashBench also boosted this little card’s cooling far beyond stock. As the water block he had was too large for the tiny 5050, a tower CPU cooler was applied to the GPU, with a fat fan attached.

The RTX 5050 came out fighting with TrashBench seeing a 3.3 GHz core clock, a 28% increase on the stock speeds. The extra shot of speed, coming from purely “raw offset” tweaking, precipitated a very respectable 17.55% gain, on average, in the handful of games TrashBench tested for this comparison. Compare that to the extra average uplift of just 3% TrashBench could achieve by tuning the old GTX 1080 Ti.

AMD’s incredibly powerful new Threadripper Pro 9995WX doesn’t come bundled with a cooler. However, Geekerwan might have the perfect answer for this processing Goliath that can casually consume over 1,000W. The Chinese TechTuber decided to remove the radiator and pump from his BMW M4 high-performance coupé, to see how it could tame the heat from Shimada Peak.

AMD’s $11,699 Threadripper Pro 9995WX is designed to deliver breakthroughs, as the “World’s Fastest Workstation Processor,” according to AMD. Checking through the full Threadripper 9000 specs table we published last week, the 9995WX is right at the top, as the newest flagship, featuring the Zen 5 architecture, a 384MB cache, and boasting 96 cores and 192 threads, running at up to 5.4GHz - given the opportunity (power and cooling) to do so.

While it has a quoted TDP of 350W, if you want to squeeze every ounce of performance out of the 9995WX, you will turn to PBO and overclocking techniques. With these performance-enhancing technologies turned up to 11, the king of the Shimada Peak processors can quickly demand 1,000W of power, or much more. That’s a lot of heat to dissipate.

BMW (and Toyota) to the rescue?

Geekerwan asserts that the BMW M4 has “the best cooling performance in the world” for a civilian vehicle (machine translated). Holding aloft the radiator, ripped from his precious M4, he highlights that the hefty rad in his hands is about 600 x 350mm, or about the size of five 360mm radiators stuck together.

Geekerwan’s BMW radiator shenanigans begin around 7 minutes. Closed captions translated to English are available.

Of course, a suitable pump is also required to make the most of the BMW radiator, and naturally, the M4’s pump was also commandeered for this Threadripper 9000 cooling task. The M4's automotive cooling pump can push 1,200 liters per hour (0.33 liters per second!).

There’s one more essential to building an automotive AIO for a CPU, and that is the fans. Geekerwan didn’t opt for the typical Arctic, Noctua, or Cooler Master spinners, though. Here’s where Toyota came in, with a Highlander 4x4’s fans borrowed for pushing air through the M4 radiator. Two of these 30cm fans were used and required 100W alone to do their duty.

Still can’t beat a flask of LN2

Tech enthusiast Matt Cole is the determined creator and maintainer of The Great MicroSD Card Survey, a significant investigative testing and benchmarking treasure trove that began in July 2023. Cole has “set up 8 machines with close to 70 card readers running around the clock… writing 101 terabytes of data per day,” in order to expose the MicroSD card diamonds in the rough and to find fake cards with inflated capacity ratings and cards that use sub-par 'fake flash.' Key features of the regularly updated 15,000-word report are its highlighting and charting of MicroSD cards alongside authenticity, performance, and endurance measurements.

The survey currently encompasses over 200 different MicroSD cards. In a blog post last year, Cole informed readers that running the tests to harvest data from these fingernail-sized slivers of flash storage resulted in over 18 petabytes of data being written to nearly 200 cards, and 51 were literally tested to destruction.

Too much information can make data almost as opaque as too little. So, thankfully, Cole has a number of quick and easy-to-glance charts to show the winners, runners-up, and losers according to the various metrics. If even that is too much for you, there is also a shortcut to the best overall picks the MicroSD card inquisitor has found.

Details of the testing methodologies, host machines, card readers, and software used are shared by Cole. Temperature measurements were also taken during the storage stress testing, but it isn’t known whether these are significant to measured performance metrics or endurance at this time.

Authenticity, 'fake flash' and 'skimpy flash'

When shopping for a MicroSD card, it is important to consider authenticity. Cole isn’t looking into counterfeit activity here; rather, he seeks to chart cards with what he calls ‘fake flash,’ where unscrupulous manufacturers and resellers will happily sell you a card with an advertised capacity that bears little relation to its storage capabilities. Cole also highlights ‘skimpy flash,’ where a ‘64GB’ card might just offer “about 61.89GB,’ for example. Skimpy flash is prevalent, even among the most trusted brands.

Performance, SDA checks

With a view to performance, Cole tested two key quality metrics. Firstly, he was interested in how well cards performed on an absolute level – raw performance. Secondly, his data checks flag whether a card’s SD Association speed rating logo is the correct one. This section also exposes the underlying data from sequential I/O test and a random I/O test results.

Endurance, Lenovo no go

Endurance is also crucial for flash storage buyers in maintaining the long-term integrity of their valuable data. How long can a user expect their MicroSD card to last under heavy read/write cycles? These tests should help deliver insight for reasoned expectations.

Cole explains that the endurance test suite will put MicroSD cards under a sustained heavy load. The most important insights we noted were that “Name-brand cards obtained from Amazon performed markedly better than name-brand cards obtained from AliExpress,” and that “fake flash performed significantly worse than authentic flash.”

So, be warned, from the best cards to the worst, the average durability almost halves, from 3,166 to 1,594 cycles. Meanwhile, if you are in a pinch and don’t have this data to refer to, it seems like a safe bet to buy a Sandisk or Kingston-branded card for an average of 4,634 and 3,555 read/write cycles before encountering their first error, respectively. Lenovo was the worst-performing well-known brand, with an average of 291 cycles before the first error occurred. Some off-brand memories lasted on average as few as 27 cycles.

Cole’s data goes far beyond this first error data, with additional charts for things like 10% failure and complete failure. However, endurance data doesn’t cover as many models as the performance charts due to its time-consuming nature.

Overall picks, weighing the data

Hong Kong-based GPU manufacturer Inno3D, which shares a parent company with Zotac, has just released benchmarks for the RTX 5050. @haruzake5719 shared the performance comparison on X, where the benchmarks showed the new GPU outperforming the RTX 3060 and running almost on par with the last-generation RTX 4060. The RTX 5050 is Nvidia’s entry-level 50-series graphics card, sporting only 8GB of slower GDDR6 and a maximum TDP of 130 watts.

This GPU launched last week and is only priced at $249, allowing budget gamers to buy into the RTX 50-series family for 1080p gaming. Although the RTX 5060 also has 8GB of VRAM, it uses GDDR7 and features more CUDA cores for better performance. It also has a higher MSRP of $299, and you can expect add-in board partners to put a premium on some variants. The RTX 5050 goes head-to-head with the Intel B580, which launched in December last year, and has the same price. However, the latter has 4GB more VRAM and a larger memory bus of 192-bit (over the RTX 5050’s 128-bit memory bus).

Inno3D’s tests show the RTX 5050 outperforming the RTX 4060 in several 3DMark tests, including Steel Nomad, Speed Way, Port Royal, Fire Strike Ultra, and Fire Strike. The latter garnered better scores in TimeSpy Extreme and TimeSpy, but its advantage sits at less than 10%. The company then ran benchmarks on four games, and although the RTX 4060 exceeded the new GPU in all titles, the FPS numbers weren’t that far off.

Today, overclocking expert and TechTuber, Roman ’der8auer’ Hartung, revealed the gaming performance that an Nvidia RTX Pro 6000 Blackwell graphics card can deliver. We’ve seen some gaming-like benchmark results for Nvidia's most powerful workstation GPU before, but der8auer also takes this $10,000 professional card though games like Cyberpunk 2077, Star Wars Outlaws, Remnant 2, and Assassin’s Creed Mirage. These scores are compared with RTX 5090 and RTX 4090 cards in the same rig. The result? Der8auer crowns the RTX 6000 “the new gaming king.”

In his 15-minute video, der8auer admits he hopes to use this RTX Pro 6000 in his personal rig, and enjoy the resulting gaming sessions. Before getting into the tests, the TechTuber reminds us that this card doesn’t have a Game Ready Driver, it only runs with its workstation driver. But this doesn’t seem to be a big issue.

The first game dissected is Cyberpunk 2077, played in 4K with max settings (no ray tracing). Der8auer’s charts show average FPS, power draw, and 1% minimum FPS for each sample. Throughout these tests, the workstation card is pitted against the RTX 5090 and 4090 - these sit among the best graphics cards of recent years.

As you can see, the RTX Pro 6000 is 14% faster than the RTX 5090, on average, and it has 13% better 1% lows in this title. However, it consumes 15% more power. If you are splurging about $10,000 on such a GPU, though, you might not care too much about the slight wrinkle in the power consumption comparison.

The gaming benchmarking showdown continued with der8auer testing the trio of graphics cards in Star Wars Outlaws (11% faster), Remnant 2 (11% faster), and Assassin’s Creed Mirage (3% faster). That last result might have been “a driver thing,” pondered the video host. It looks like the previously seen 3DMark tests actually gave quite a good window onto the performance here.

If you haven’t already done so, you can watch the full der8auer video above, where lots of other background details are shared. You can also see a fun comparative Cyberpunk 2077 benchmark with its most extreme settings (path tracing) and MFG turned on.

Recap of the RTX Pro 6000 specs

You can read our previous coverage for more in the way of tech specs, but it is worth reminding readers that the RTX Pro 6000 is based on the same GB202 GPU as the RTX 5090. However, the workstation card benefits from 24,064 shaders (rather than 21,760) as well as more TMUs, ROPs, Tensor cores and RT cores. Nevertheless, the memory configuration is probably its standout feature in the comparison with the RTX Pro 6000 packing 96GB of GDDR7 on its 512-bit bus. A consumer-grade RTX 5090 packs ‘just’ 32GB.

Coil whine and loud fans

With its FE-style cooler design the RTX Pro 6000 is nice and compact, but with it easily reaching 600W power consumption, there are a few comments made about the noise of this graphics card. Wryly, der8auer says that the fans ramp up so fast and so loud that you can’t really hear the coil whine. Despite the fans, “This is the worst coil whine I’ve ever heard, it’s insanely loud,” he observes.

Can’t ignore the pricing

We're seeing traces of AMD's more affordable Strix Halo options, with the octa-core Ryzen AI Max 385 marking its debut on Geekbench. Powering the HP ZBook Ultra G1a, Strix Halo appears poised to expand into the mainstream segment with eight and six-core variations. Performance indicators in Geekbench aren't as impressive as we expected, though we shouldn't consider them as definitive due to this being a leak, typical run-to-run variations, and laptop power settings.

CES marked the debut of AMD's flagship APU family this generation, Strix Halo, under the Ryzen AI 300 Max line, for AI/ML developers and gamers. AMD launched four different variants, ranging from 6-16 CPU cores and 16-40 GPU cores, catering to a wide range of consumers. Most of the first and second quarters have been occupied by the premier Ryzen AI Max+ 395 and 390, with mini-PCs costing up to $2,000. Meanwhile, laptops, limited to the HP Zbook Ultra G1a and Asus ROG Flow Z13, exceed the $5,000 price point, fully decked out with 128GB of LPDDR5x RAM.

The Ryzen AI Max 385 sports eight cores and sixteen threads based on the Zen 5 architecture, alongside a 32 Compute Unit equipped integrated GPU (Radeon 8050S) based on RDNA 3.5 and a 50 TOPS capable NPU. The CPU is advertised to reach 5 GHz, only 100 MHz below the flagship model. This suggests similar single-core performance, but the actual results indicate otherwise.

Earlier this month, reports suggested that Nvidia has added SK hynix as a GDDR7 supplier for its RTX 50 series GPUs, after previously only sourcing from Samsung. On Chiphell, a user recently got their hands on a Gigabyte Gaming OC RTX 5070 Ti with SK hynix memory and found its overclocking potential similar to what we've seen from Samsung, via UNIKO's Hardware. While thermal and efficiency data are missing, this is nonetheless a positive indication.

Each generation, Nvidia casts a wide net to broaden its memory supplier base, ensuring multiple options. This is a common approach with many products, especially SSDs, many of which often ship with revised controllers or NAND flash modules later in their lifecycle. This shouldn't impact the average consumer as Nvidia likely validates all memory chips to run at a minimum of 28 Gbps (or 30 Gbps for the RTX 5080), at defined power, voltage, and temperature settings.

A user at Chiphell secured an RTX 5070 Ti from Gigabyte (SK hynix memory) earlier this week. Several days later, they attempted to increase the power limits by flashing a BIOS that was designated for the Aorus Master RTX 5070 Ti (Samsung memory), resultantly bricking the card. Fortunately, the damage wasn't permanent as their GPU was equipped with a dual BIOS, allowing them to switch to the secondary firmware.

Either way, to test the limits of their GPU's memory, the user achieved a memory clock of 2,125 MHz (34 Gbps) with overclocking, aligning with the majority of Samsung overclocks we've witnessed. This tells us that GDDR7 modules from SK hynix are just as capable as ones from Samsung when it comes to overclocking - with this sample.

GPU Benchmarks Introduction

Tom's Hardware exhaustively benchmarks every GPU to find out which are worthy of our list of the best graphics cards. Our GPU benchmarks hierarchy ranks current and previous generation graphics cards by performance. Whether it's playing games, running artificial intelligence workloads, or doing professional video editing, your graphics card typically plays the biggest role in determining performance — even the best CPUs for gaming take a secondary role.

Our 2026 GPU Hierarchy testing spans three generations of Nvidia and AMD graphics cards, as well as Intel's Arc B-series GPUs.

Our testing has been made easier by the fact that no truly new gaming GPUs have been introduced in almost a year. If you haven't already upgraded your graphics card after the GeForce RTX 50-series and Radeon RX 9000-series launches in 2025, well, you're still looking at the exact same products now.

AMD did make its formerly China-only Radeon RX 9070 GRE available globally after Computex 2026, but in our review, we found that $549 product to be too expensive given the level of performance it delivers and the compromises made to hit its price point. Check out that coverage for all the details.

Most of the products we recommend remain at elevated prices compared to their MSRPs, but this is just life in mid-2026. It's admittedly cold comfort, but unless you're shopping for an RTX 5090, graphics card prices haven't risen much more than they already did earlier this year.

Compared to the doubling or tripling of prices we've seen for RAM kits and SSDs in 2026 versus last year, a GPU upgrade remains a relatively affordable (and self-contained) option, either as a boost for an existing PC or part of an all-new parts list.

Let's dive into our ranking of GPUs past and present so you can figure out how all those cards stack up.

Prime Day exceptional graphics card deals

These are a few of the standout deals from Amazon's 2026 Prime Day event, which is currently taking place. Be sure to check out our full list of the best Prime Day graphics card deals, too. The GPU Benchmarks Hierarchy continues below.

Our full GPU hierarchy using traditional rendering (aka rasterization) comes first, and below that we have our separate ray tracing hierarchy. We've also mashed up these results into one overarching ranking for the complete pictures.

The results are all collected at native resolution, without enabling DLSS, FSR, or XeSS upscaling or frame generation.

As a brief refresher of the cards in this version of our hierarchy, Nvidia's cutting-edge Blackwell architecture and its DLSS 4 suite of tech upgrades the quality of both upscaling and ray reconstruction on RTX 50-series GPUs. It also adds Multi Frame Generation support. As of mid-2026, MFG can add anywhere from one to five AI-generated frames in between natively rendered ones.

Even if you're not into framegen, DLSS 4.5 upscaling and its transformer neural network architecture can offer a big boost in image quality at the same output resolution compared to earlier DLSS versions. That tech can benefit all GeForce RTX GPUs going back to the RTX 20-series, albeit at a heavier performance cost on older hardware compared to earlier DLSS versions.

Nvidia's Ada Lovelace architecture powers its prior-gen RTX 40-series cards. Ada introduced DLSS Frame Generation, which can double output frame rates in supported games. Ada cards don't benefit from MFG, though.

AMD's Radeon RX 9000 series cards, powered by its latest RDNA 4 architecture, get a big boost to both ray-tracing and AI capabilities with dedicated RT and matrix math accelerators. AMD uses those capabilities to enable its FSR 4 upscaler and its much-improved image quality in a small but growing range of titles, whether through native support or with driver-level overrides.

The FSR Redstone update last year brought ML Frame Generation to RX 9000-series cards. Like FSR 4.x upscaling, ML Frame Generation can be directly integrated in games or enabled through a driver override.

Meanwhile, the last-gen RDNA 3 architecture powers seven RX 7000-series seven desktop cards. Until very recently, RX 7000-series cards couldn't officially run FSR 4 upscaling, but an official version of that tech will be made available for those cards in July 2026.

Intel's Battlemage-powered Arc B580 and Arc B570 offer major improvements in performance and efficiency compared to its Alchemist architecture. Battlemage only serves the entry-level end of the graphics card market today. Intel introduced a larger Battlemage chip in the form of the Arc Pro B70 earlier this year, but that product is intended for AI and professional visualization, and it's priced like it.

On page two, you'll find our 2025-2026 test data for posterity. Page three has our 2024–2022 benchmark suite, which covers previous-generation GPUs running an older test suite and a Core i9-12900K. Page four has an even older 2020–2021 test suite with only raster games, running on a Core i9-9900K testbed. The legacy tables are no longer being actively updated. There's also the "Ancient Legacy GPU hierarchy" (without benchmarks, sorted by theoretical performance) for reference purposes.

The following tables sort everything solely by our performance-based GPU benchmarks, from fastest to slowest. Factors including real-world pricing, graphics card power consumption, overall efficiency, and features aren't factored into the rankings here. The latest results use an AMD Ryzen 7 9800X3D testbed. Here are the tables and benchmark results — rasterization games first, then ray tracing games, and finally the content creation results.

GPU Benchmarks Hierarchy 2026: The tests

For our latest GPU benchmarks, we test every card at a mix of high and ultra settings, depending on the game. We test across three resolutions: 1080p, 1440p, and 4K.

All the scores are expressed as percentages relative to the top-ranking 1080p ultra card, which is of course the GeForce RTX 5090.

Our 2026 test suite comprises the following games. 11 are raster titles, and of those, four have RT support. We test another four games that either require RT to run (DOOM: The Dark Ages and Indiana Jones and the Great Circle) or look best with RT (Assassin's Creed Shadows and Grand Theft Auto V Enhanced.)

• Black Myth Wukong (+RT)
• Alan Wake II (+RT)
• Apex Legends
• Counter-Strike 2
• Fortnite
• Arc Raiders
• Stalker 2
• DOOM: The Dark Ages (RT required)
• Grand Theft Auto V Enhanced (RT only)
• Marvel's Spider-Man 2 (+RT)
• Indiana Jones and the Great Circle (RT required)
• Marvel Rivals
• Assassin's Creed Shadows (RT only)
• Cyberpunk 2077 (+RT)
• Clair Obscur: Expedition 33

GPU Benchmarks Hierarchy 2026: Raster gaming

The FPS score is the geometric mean (equal weighting) of all 11 games. Note that the specifications column links directly to our original review for the various GPUs.

Rasterization GPU Benchmarks, Key Findings

• Unsurprisingly, the RTX 5090 takes the top spot across the board, but prices remain stratospheric in mid-2026. It's also difficult to fully engage the GB202 GPU for pure raster gaming unless you're playing at 4K.
• Among current cards, the RX 9060 XT 8GB and RTX 5060 duke it out for the best performance per dollar at 1080p.
• The Radeon RX 9070 is the 1440p raster value champ. The recently introduced RX 9070 GRE sits just behind, and the RTX 5070 is a somewhat distant third.
• At 4K, the RX 9070 XT is the cheapest way to get into 4K 60 FPS native gaming. But the RX 9070 and RTX 5070 are also strong options among current GPUs if you're willing to enable a dash of upscaling.

Overall, if you're only interested in native raster gaming, Radeons are a strong choice in 2026. But Nvidia offers superior DLSS 4.5 upscaling and Multi Frame Generation support across all RTX 50-series cards, which makes matching your GPU's performance to your monitor's refresh rate a snap.

FSR 4.x upscaling isn't on par with DLSS 4.5 yet, and FSR ML Frame Generation is limited to a 2x framerate boost where it's available.

Rasterization GPU Benchmarks Rankings 2026

Remember that we're not including any upscaling or framegen results in the above table. DLSS, FSR, and XeSS offer differing image quality, and we want to keep things directly comparable.

Don't buy any of the cards at the top of our hierarchy without a high-refresh-rate 2560x1440 or 4K monitor to match. Even one of our best gaming CPUs, like the Ryzen 7 9800X3D, can only do so much when a powerful card becomes CPU-bound.

In general, if you're still running a 60 Hz fixed-refresh-rate monitor, it's entirely possible that you're not seeing all the frames your graphics card can generate. Practically any current-gen graphics card from the RTX 5060 on up is good enough for high-refresh-rate gaming at 1080p in 2026 even without upscaling or framegen, and the RTX 5070 or RX 9070 will give you the same experience at 1440p.

In this era of upscaling and framegen, a high-resolution monitor is no longer an obstacle to the best gaming experience. Upscalers work best with higher-than-1080p output resolutions, and the advent of the DLSS 4.5 upscaling model especially means that you can get both smooth performance and crisp image quality from surprisingly modest graphics cards.

But if you don't have a high-refresh-rate 1440p or 4K monitor to begin with, you can't take full advantage of the free boost to both performance and image quality that AI-powered upscaling offers, nor can you enjoy the full smoothness boost of framegen. Check out our list of the best gaming monitors for a potential upgrade if you're still rocking a 1080p 60 Hz display from the 2010s.

On to our 2026 raster results. Among currently available graphics cards, Blackwell rules the top three spots. The RX 9070 XT ends up fourth, just behind the potent (but much more expensive, as of June 2026) RTX 5070 Ti.

Moving into the midrange, the RX 9070's strong raster performance gives it the edge over the RTX 5070, but it's close. And the 5070 has the full arsenal of DLSS 4.5 features at its disposal in virtually every modern game.

Given the image quality and smoothness advantage of DLSS 4.5 upscaling and MFG, the 5070 got the nod in our most recent round of best graphics card picks after our 2026 retests.

The recently introduced RX 9070 GRE lands between the 5070 and the lower-end 5060 Ti and 9060 XT, a position for which Nvidia has no current-gen answer (and no easy path to one.) But AMD may have priced the GRE too high at $549, and we're already seeing big price drops on that product that might make it uniquely appealing in a market where the RTX 5060 Ti is its direct price competition.

The formerly midrange RTX 5060 Ti 16GB is hopelessly overpriced for the performance it offers in mid-2026, and the formerly somewhat affordable 8GB model is also selling for elevated prices now, making the entire 5060 Ti family impossible to recommend.

The RX 9060 XT 16GB holds down the entry-midrange 16GB position by itself, and its $459 price tag isn't so absurd as to make it un-recommendable. But as RX 9070 GREs start going on sale for $500-ish, the step up is going to look mighty tempting.

As we move further down the stack, the $300-ish Arc B580 represents Intel's best card right now, but it remains very much a budget part in mid-2026, trading blows with the RTX 3060 Ti and RTX 4060. The only reason it lands as high as it does in our results is that its 12GB of VRAM prevents its performance from totally plummeting at 4K, as it does for those older 8GB cards.

The RTX 5060 is 24% faster than the Arc B580 at 1080p and 17% faster at 1440p, and its prices are proportionately higher. The scarce RX 9060 XT 8GB is a whole 30% faster at 1080p and 23% faster at 1440p, for similarly more money.

We don't think the potential longevity afforded by the B580's 12GB of VRAM is worth the massive performance tradeoff versus current-gen 8GB GPUs. Save up a few more bucks and grab an RTX 5060 or RX 9060 8GB instead (and enjoy the more widely adopted DLSS 4.5 or FSR 4 versus XeSS while you're at it).

The $250 Arc B570 is still a good value in mid-2026, and as the only modern, decent graphics card available for (well) under $300, it's still worth a look if you can't stretch to an RTX 5050. But recent software ills that took months to fix have led us to consider it more of a gamble than a product worthy of a recommendation, even though those issues did eventually get corrected.

Ray Tracing GPU Benchmarks 2026

For 2026, we're testing a range of RT titles that present a progressively more difficult workload to the GPUs under test.

Grand Theft Auto V Enhanced, DOOM: The Dark Ages, and Indiana Jones and the Great Circle are all relatively lightweight RT titles that will run on a wide range of RT-ready hardware. Spider-Man 2, Cyberpunk 2077, and Assassin's Creed Shadows represent greater challenges to compute resources, VRAM, or both. And Black Myth Wukong and Alan Wake II still bring even the most powerful graphics cards to their knees at native resolutions.

Our results for those more demanding games represent more of a jumping-off point for upscaling and framegen rather than native performance. But lighter RT titles mean that you can explore the differences in image quality that ray tracing offers without crushing your GPU flat.

In any case, graphics cards at the top of our RT tests are ready for practically any RT game. But as you'll see, that's still quite the high bar to clear in 2026.

Ray Tracing GPU Benchmarks, Key Findings and Notes

• The RTX 5090 really comes into its own with RT enabled, as expected.
• AMD's RX 9070 XT outperforms the RX 7900 XTX in our RT tests, showing that the ray-tracing improvements in RDNA 4 deliver.
• Even with those improvements, the RX 9070 XT can only manage a tie with the RTX 4070 Ti with RT in play, coming in eighth place overall.
• Maintaining 60 FPS in RT titles (as a good foundation for upscaling and framegen enhancements) is a very tall order. Among recent products, you'll want:
  • an RTX 5060 Ti 16GB, RTX 5070, or RX 9070 GRE to handle ~60 FPS RT at 1080p without running out of VRAM
  • An RTX 4070, an RTX 5070 or RX 9070 at 1440p, or an RX 9070 XT or RTX 5070 Ti for the best possible experience
  • An RTX 5080, RTX 4090, or RTX 5090 at 4K

Ray Tracing GPU Benchmarks Rankings 2026

The RTX 5090 and RTX 4090 lead our RT results. But if you want to enjoy high-end RT without a 4090- or 5090-sized bankroll, the RTX 5070 Ti proves itself as the last card with a reasonable price-to-performance ratio before things get crazy.

The 5070 Ti leads the RX 9070 XT by a wide margin in our RT tests, and it's anywhere from 10%-15% behind the absurdly overpriced RTX 5080 across the board. But if you want the absolute best RT performance possible without stepping up to a 4090 or 5090, the extra cash for a 5080 is your best — and only — option.

That said, you should really explore DLSS 4.5 upscaling and MFG on the 5070 Ti before spending big to step up to a 5080.

The RX 9070 XT remains AMD's fastest RT GPU, outpacing the 7900 XTX across all tested resolutions. It leads the RDNA 3 halo card by about 10% across the board, which is an impressive generational leap considering its 33% deficit in CUs versus the XTX.

But the 9070 XT can only trade blows with the RTX 4070 Ti and RTX 5070 with Nvidia competition in the picture. It beats out the 5070 at 1080p and 1440p, but DLSS 4.5 upscaling and MFG give Blackwell gamers more flexible options for boosting performance with only minimal impact to image quality versus both RDNA 4 (and Ada).

Of course, the RX 9070 XT has FSR 4 upscaling and framegen at its disposal in compatible titles, but support for those features isn't as widespread as DLSS.

Meanwhile, the RX 9070 comes in slightly behind the RTX 5070 at 1080p and 1440p for RT. Even though its 16GB of VRAM prevents the 9070's 4K RT performance from completely nosediving, the average frame rate at that resolution isn't high enough to make it a 4K RT choice.

At 1080p, the RX 9070 GRE presents an intriguing RT option for Radeon gamers, delivering a near-60-FPS average baseline that would make a great FSR 4 foundation. But its 12GB of VRAM quickly becomes a liability at 1440p versus the 9070, and 4K is out of the question.

Overall, our demanding 2026 RT test suite shows that the bar for entry to a good RT experience remains high, especially as elevated prices on the RTX 5060 Ti 16GB mean that the first "real" RT-ready Blackwell card at 1080p and up is the RTX 5070. And on the Radeon side, you really want an RX 9070 GRE or RX 9070 for the best experience.

Test Setup and Hardware for GPU Benchmarks

We've used several different PCs for our testing over the years. The latest update switches to an AMD Zen 5 processor: the unparalleled Ryzen 7 9800X3D and its 3D V-Cache-enhanced performance.

Here are the specifications for our latest GPU test PC.

Tom's Hardware 2026 GPU Benchmarks Testbed

CPU: AMD Ryzen 7 9800X3D
CPU Cooler: Thermalright Phantom Spirit 120 SE
Memory: G.Skill Trident Z5 32GB (2x16GB) DDR5-6000 CL30
Motherboard: Asus TUF Gaming X670E-Plus Wifi
SSD: Inland Performance Plus 4TB
PSU: MSI MPG Ai1600TS 1600W

We test across the three most common gaming resolutions, 1080p, 1440p, and 4K, using a mix of high and ultra settings, depending on the title. Where possible, we use 'reference' cards for all of these tests, like Nvidia's Founders Edition models and AMD's reference designs. Most midrange and lower GPUs don't get reference models, however, and in some cases we only have factory-overclocked cards for testing. We do our best to select cards that are close to the reference specs in such cases.

For each graphics card, we follow the same testing procedure. We run one pass of each benchmark to "warm up" the GPU after launching the game, then perform our actual test runs across each resolution.

We carefully review our test data and check for anomalies. For example, we always expect the RTX 5080 to be faster than the RTX 5070 Ti. If it's not, and we're not in a CPU limited situation, we'll recheck both cards to ensure that our standings our accurate. We also check and retest in cases of subtler issues, as when a transient hitch or frame-time spike causes a large dip in 1% low FPS.

Due to the length of time required for testing each GPU, updated drivers and game patches inevitably come out that can impact performance. We periodically retest a few sample cards to verify our results are still valid, and if not, we go through and retest the affected game(s) and GPU(s). We may also add games to our test suite over time, if one comes out that is popular and conducive to testing. See what makes a good game benchmark for our selection criteria.

GPU Benchmarks: Individual Game Charts

The above tables provide a summary of performance, but for those that want to see the individual game charts, for both the standard and ray tracing test suites, we've got those as well. These charts only cover current-gen GPUs for readability.

These charts are up to date as of June 2026.

GPU Benchmarks — 1080p

GPU Benchmarks — 1440p

GPU Benchmarks — 4K

GPU Benchmarks — Power, Clocks, and Temperatures

Most of our discussion has focused on performance, but for those interested in power and other aspects of the GPUs, here are the appropriate charts. We'll run these from highest to lowest settings, as 4K ultra tends to be the most strenuous workload on most of these GPUs.

If you're looking for the legacy GPU hierarchy, head over to page two! We moved it to a separate page to help improve load times in our CMS as well as for the main website. And if you're looking to comment on the GPU benchmarks hierarchy, head over to our forums and join the discussion!

Choosing a Graphics Card

Choosing a Graphics Card

Choosing a Graphics Card

Which graphics card do you need? To help you decide, we created this GPU benchmarks hierarchy, comprising 48 GPUs from three generations of hardware for Nvidia and AMD and also including Intel's Battlemage cards. Not surprisingly, the fastest cards are using the latest GPU architectures, though they're not always a major upgrade over the prior generation.

Of course, it's not just about playing games. Many applications use the GPU for other work. But a good graphics card for gaming will typically do equally well in complex GPU computational workloads. Buy one of the top cards and you can run games at high resolutions and frame rates with the effects turned all the way up, and you'll be able to do content creation work as needed. Drop down to the middle and lower portions of the list and you'll need to start dialing down the settings to get acceptable performance in regular gameplay and GPU benchmarks.

If your main goal is gaming, you can't forget about the CPU. Getting the best possible gaming GPU won't help you much if your CPU is underpowered or many years out of date. So be sure to check out the Best CPUs for gaming page, as well as our CPU Benchmarks Hierarchy to make sure you have the right CPU for the level of gaming performance you're looking to achieve.

And don't forget about your monitor. Be sure to get a display whose refresh rate range matches the average frame rates of the graphics card that you want in the games that you play. Check out our list of the best gaming monitors for a starting point.

GPU Benchmarks from 2022–2024

Our 'legacy' GPU benchmarks used different hardware. Here are the details for the 2022–2024 configuration that used an Alder Lake 12900K CPU.

Tom's Hardware 2022–2024 GPU Testbed

Intel Core i9-12900K
MSI Pro Z690-A WiFi DDR4
Corsair 2x16GB DDR4-3600 CL16
Crucial P5 Plus 2TB
Cooler Master MWE 1250 V2 Gold
Cooler Master PL360 Flux
Cooler Master HAF500
Windows 11 Pro 64-bit

We have upgraded our test system and changed our test suite for 2025, effectively resetting our benchmarking and rankings. However, the old data — collected using a Core i9-12900K PC — remains valid. We aren't testing new GPUs with the old testbed, as that would double the time required for what is already a lengthy process, but we felt it would be helpful to some to keep the final data available.

We also had an interim system that never quite got wrapped up, which had a Core i9-13900K Raptor Lake CPU. Our original CPU was one of the impacted units that failed over time, which explains in retrospect why so much of the data felt a bit questionable. Thankfully, the new Ryzen 7 9800X3D system seems to be running just fine. Here are the 12900K results.

For these GPU benchmarks, we tested nearly every GPU released between 2016 and 2024, plus a few extras. All graphics cards were tested at 1080p medium and 1080p ultra, and we sorted the table by the 1080p ultra results. Where it made sense, we also tested at 1440p ultra and 4K ultra. All the scores are scaled relative to the top-ranking 1080p ultra card of the time, which in our suite is the RTX 4090 — especially at 4K and 1440p.

The above summary charts show the relative performance of the cards we've tested across the past several generations of hardware at 1080p ultra — swipe through the above gallery if you want to see the 1080p medium, 1440p, and 4K ultra images. There are a few missing options (e.g., the GT 1030, RX 550, and several Titan cards), but otherwise it's basically complete. Note that we also have data in the table below for some of the other older GPUs.

The eight games used for our standard GPU benchmarks hierarchy are Borderlands 3 (DX12), Far Cry 6 (DX12), Flight Simulator (DX11 Nvidia, DX12 AMD/Intel), Forza Horizon 5 (DX12), Horizon Zero Dawn (DX12), Red Dead Redemption 2 (Vulkan), Total War Warhammer 3 (DX11), and Watch Dogs Legion (DX12). The fps score is the geometric mean (equal weighting) of the eight games. Note that the specifications column links directly to our original review for the various GPUs.

*: GPU couldn't run all tests, so the overall score is slightly skewed at 1080p ultra.

While the RTX 4090 does technically take first place at 1080p ultra, it's the 1440p and especially 4K numbers that impress. It's less than 2% faster than the RTX 4080 Super at 1080p ultra, but that increases to 9% at 1440p and then 25% at 4K. Also note that the fps numbers in our table incorporate both the average and minimum fps into a single score — with the average given more weight than the 1% low fps.

Again, keep in mind that we're not including any ray tracing or DLSS results in the above table, as we use the same test suite with the same settings on all current and previous generation graphics cards. Since only RTX cards support DLSS (and RTX 40-series if you want DLSS 3), that would drastically limit which cards we could directly compare. You can see DLSS 2/3 and FSR 2 upscaling results in our RTX 4070 review if you want to check out how the various upscaling modes can help.

The RTX 4090 comes at a steep price, though on paper it's not that much worse than the previous generation RTX 3090. In fact, we'd say it's a lot better in some respects, as the 3090 was only a minor improvement in performance compared to the 3080 at the time of launch, but with more than double the VRAM. Nvidia pulled out all the stops with the 4090, increasing the core counts, clock speeds, and power limits to push it beyond all contenders. There are two problems with the 4090, however: It's not available at MSRP any longer, due to demand from the AI sector — it often costs $2,000 or more — and there are still concerns with pulling 450W of power over the 16-pin connector.

Stepping down from the RTX 4090, the RTX 4080 Super and RX 7900 XTX trade blows at higher resolutions, while CPU bottlenecks come into play at 1080p. We'll be switching our testbed in the near future, with the current results from our 13900K testing in the charts at the bottom of the page.

Outside of the latest releases from AMD and Nvidia, the RX 6000- and RTX 30-series chips still perform reasonably well and if you're using such a card, there may not be any need to upgrade at present. Intel's Arc GPUs also fall into this category and are something of a wild card.

We've been testing and retesting GPUs periodically, and the Arc chips running the latest drivers now complete all of our benchmarks without any major anomalies. (Minecraft was previously a problem, though Intel has finally sorted that out.) They're not great on efficiency, but overall performance and pricing for the A750 is quite good.

Turning to the previous generation GPUs, the RTX 20-series and GTX 16-series chips end up scattered throughout the results, along with the RX 5000-series. The general rule of thumb is that you get one or two "model upgrades" with the newer architectures, so for example the RTX 2080 Super comes in just below the RTX 3060 Ti, while the RX 5700 XT basically matches the newer and less expensive RX 6600 XT.

Go back far enough and you can see how modern games at ultra settings severely punish cards that don't have more than 4GB VRAM. We've been saying for a few years now that 4GB was just scraping by, and these days we'd avoid buying anything with less than 8GB of VRAM — 12GB or more is the minimum we'd want with a mainstream GPU, and 16GB or more for high-end and above. Old cards like the GTX 1060 3GB and GTX 1050 actually failed to run some of our tests, which skews their results a bit, even though they do better at 1080p medium.

Now let's switch over to the ray tracing hierarchy.

Ray Tracing GPU Benchmarks 2022–2024

Enabling ray tracing, particularly with demanding games like many of those we're using in our DXR test suite, can cause framerates to drop off a cliff. We're testing with "medium" and "ultra" ray tracing settings. Medium generally means using the medium graphics preset but turning on ray tracing effects (set to "medium" if that's an option; otherwise, "on"), while ultra turns on all of the RT options at more or less maximum quality.

Because ray tracing is so much more demanding, we're sorting these results by the 1080p medium scores. That's also because the RX 6500 XT, RX 6400, and Arc A380 basically can't handle ray tracing even at these settings, and testing at anything more than 1080p medium would be fruitless.

The five ray tracing games used are Bright Memory Infinite, Control Ultimate Edition, Cyberpunk 2077, Metro Exodus Enhanced, and Minecraft — all of these use the DirectX 12 / DX12 Ultimate API. The fps score is the geometric mean (equal weighting) of the five games, and the percentage is scaled relative to the fastest GPU in the list, which again is the GeForce RTX 4090.

If you want to see what the future may hold with ray tracing, check out our Alan Wake 2 benchmarks where the full path tracing barely manages playable performance even with upscaling on non-Nvidia GPUs. However, and this is a critical point, the number of games where RT truly makes a striking difference in visuals is extremely limited. For most games, we still feel running pure rasterization rendering makes more sense.

If you felt the RTX 4090 performance was impressive at 4K in our standard test suite, just take a look at the results with ray tracing. Nvidia put even more ray tracing enhancements into the Ada Lovelace architecture, and those start to show up here. There are still further potential performance improvements for ray tracing with SER, OMM, and DMM — not to mention DLSS 3, though that ends up being a bit of a mixed bag, since the generated frames don't include new user input and add latency.

If you want a real kick in the pants, we also ran many of the faster ray tracing GPUs through Cyberpunk 2077's RT Overdrive mode, which implements full "path tracing" (full ray tracing, without any rasterization) — as well as Alan Wake 2, which uses path tracing at higher settings, and Black Myth: Wukong that supports full ray tracing. Those games provide a glimpse of how future games could behave, and why upscaling and AI techniques like frame generation are here to stay.

Even at 1080p medium, a relatively tame setting for DXR (DirectX Raytracing), the RTX 4090 roars past all contenders and leads the previous generation RTX 3090 Ti by 41%. At 1080p ultra, the lead grows to 53%, and it's nearly 64% at 1440p. Nvidia made claims before the RTX 4090 launch that it was "2x to 4x faster than the RTX 3090 Ti" — factoring in DLSS 3's Frame Generation technology — but even without DLSS 3, the 4090 is 72% faster than the 3090 Ti at 4K.

AMD continued to relegate DXR and ray tracing to secondary status, focusing more on improving rasterization performance — and on reducing manufacturing costs through the use of chiplets on the new RDNA 3 GPUs. As such, the ray tracing performance from AMD isn't particularly impressive. The top RX 7900 XTX basically matches Nvidia's previous generation RTX 3080 12GB, which puts it barely ahead of the RTX 4070 — and that's not even in all DXR games. There are some minor improvements for RT performance in RDNA 3, though, as the 7800 XT for example ends up basically tied with the RX 6800 XT in rasterization performance but is 10% faster in DXR performance.

Intel's Arc A7-series parts show a decent blend of performance in general, with the A750 coming in ahead of the RTX 3060 overall. With the latest drivers (and with vsync forced off in the options.txt file), Minecraft performance also looks much more in line with the other Arc DXR results.

You can also see what DLSS Quality mode did for performance in DXR games on the RTX 4090 in our review, but the short summary is that it boosted performance by 78% at 4K ultra. DLSS 3 frame generation improved framerates another 30% to 100% in our testing, though we recommend exercising (extreme) caution when looking at FPS with the feature enabled. It can boost framerates in benchmarks, but when actually playing games it often doesn't feel much faster than without the feature.

Overall, with DLSS 2, the 4090 in our ray tracing test suite is nearly four times as fast as AMD's RX 7900 XTX. Ouch. AMD's FSR 2 and FSR 3 can help as well, and AMD continues to work on increasing the rate of adoption, but it still trails DLSS both in the number of games supported and in the overall image quality. Only two of the games in our DXR suite have FSR2 support. By comparison, all of the DXR games we're testing support DLSS2 — and one also supports DLSS3.

Without FSR2, AMD's fastest GPUs can only clear 60 fps at 1080p ultra, while remaining decently playable at 1440p with 40–50 fps on average. But native 4K DXR remains out of reach for just about every GPU, with only the 3090 Ti and above breaking the 30 fps mark on the composite score — and a couple of games still come up short on the 3090 Ti.

AMD also has FSR 3 frame generation. Like DLSS3, it adds latency, and AMD requires the integration of Anti-Lag+ support in games that use FSR 3. But Anti-Lag+ only works with AMD GPUs, which means non-AMD cards will likely incur a larger latency penalty. We've tested it in Avatar: Frontiers of Pandora and found it worked pretty well, but that was not the case in Forspoken and Immortals of Aveum. It has since gained a lot more traction, though quality and latency remain quite variable — it can look and run well in one game, and then fall flat in another.

The midrange GPUs like the RTX 3070 and RX 6700 XT basically manage 1080p ultra and not much more, while the bottom tier of DXR-capable GPUs barely manage 1080p medium — and the RX 6500 XT can't even do that, with single digit framerates in most of our test suite, and one game that wouldn't even work at our chosen "medium" settings. (Control requires at least 6GB VRAM to let you enable ray tracing.)

Intel's Arc A380 ends up just ahead of the RX 6500 XT in ray tracing performance, which is interesting considering it only has 8 RTUs going up against AMD's 16 Ray Accelerators. Intel posted a deep dive into its ray tracing hardware, and Arc seems reasonably impressive, except for the fact that the number of RTUs severely limits performance. The top-end A770 still only has 32 RTUs, which proves sufficient for it to pull ahead (barely) of the RTX 3060 in DXR testing, but it can't go much further than that. Arc A750 and above also ends up ahead of AMD's RX 6750 XT in DXR performance, showing just how poor AMD's RDNA 2 hardware is when it comes to ray tracing.

It's also interesting to look at the generational performance of Nvidia's RTX cards. The slowest 20-series GPU, the RTX 2060, still outperforms the newer RTX 3050 by a bit, but the fastest RTX 2080 Ti comes in a bit behind the RTX 3070. Where the 2080 Ti basically doubled the performance of the 2060, the 3090 delivers about triple the performance of the 3050.

Test System and How We Test for GPU Benchmarks

For each graphics card, we follow the same testing procedure. We run one pass of each benchmark to "warm up" the GPU after launching the game, then run at least two passes at each setting/resolution combination. If the two runs are basically identical (within 0.5% or less difference), we use the faster of the two runs. If there's more than a small difference, we run the test at least twice more to determine what "normal" performance is supposed to be.

We also look at all the data and check for anomalies, so for example RTX 3070 Ti, RTX 3070, and RTX 3060 Ti all generally going to perform within a narrow range — 3070 Ti is about 5% faster than 3070, which is about 5% faster than 3060 Ti. If we see games where there are clear outliers (i.e. performance is more than 10% higher for the cards just mentioned), we'll go back and retest whatever cards are showing the anomaly and figure out what the "correct" result would be.

Due to the length of time required for testing each GPU, updated drivers and game patches inevitably will come out that can impact performance. We periodically retest a few sample cards to verify our results are still valid, and if not, we go through and retest the affected game(s) and GPU(s). We may also add games to our test suite over the coming year, if one comes out that is popular and conducive to testing — see our what makes a good game benchmark for our selection criteria.

GPU Benchmarks: Individual Game Charts

The above tables provide a summary of performance, but for those that want to see the individual game charts, for both the standard and ray tracing test suites, we've got those as well. We're only including more recent GPUs in these charts, as otherwise things get very messy. These are also using a newer test suite and a 13900K CPU, which changes the performance slightly from the above table, simply because our newest tests are more relevant (but haven't been run on a lot of the older GPUs shown in the tables).

These charts were up to date as of November 11, 2024.

GPU Benchmarks — 1080p Medium

GPU Benchmarks — 1080p Ultra

GPU Benchmarks — 1440p Ultra

GPU Benchmarks — 4K Ultra

GPU Benchmarks — Power, Clocks, and Temperatures

Most of our discussion has focused on performance, but for those interested in power and other aspects of the GPUs, here are the appropriate charts.

Our even older 2020–2021 'legacy' GPU benchmarks used another set of hardware with different games. Here are the details for the Coffee Lake 9900K and Z390 PC.

Tom's Hardware 2020–2021 GPU Testbed

Intel Core i9-9900K
Corsair H150i Pro RGB
MSI MEG Z390 Ace
Corsair 2x16GB DDR4-3200
XPG SX8200 Pro 2TB
Windows 10 Pro (21H1)

The results below combine results from nine games with six resolution and setting combinations. All of the scores are combined (via a geometric mean calculation) into a single overall result, which tends to penalize the fastest and slowest GPUs — CPU bottlenecks come into play at 1080p medium, while VRAM limitations can kill performance at 4K ultra.

These results have not been updated since early 2022, when we added the RTX 3050 and RX 6500 XT to the list. We won't be adding future GPUs to this table, so there's no RTX 40-series, RX 7000-series, Arc, 3090 Ti, 6950 XT, 6750 XT, or 6650 XT, but it does help to provide a look at a slightly less demanding suite of games, where 6GB or more VRAM isn't generally required at 1080p ultra settings. You can use these older results to help inform your purchase decisions, if you don't typically run the latest games at maxed out settings.

Ancient Legacy GPU Hierarchy

Below is our legacy desktop GPU hierarchy dating back to the late 1990s. We have not tested most of these cards in many years, driver support has ended on most models, and the relative rankings are pretty coarse. Note that we also don't factor in memory bandwidth or features like AMD's Infinity Cache or Nvidia's larger L2 cache on Ada Lovelace. The list below is mostly intended to show relative performance between architectures from a similar time period.

We sorted the table by the theoretical GFLOPS, though on architectures that don't support unified shaders, we only have data for "Gops/s" (giga operations per second). That's GeForce 7 and Radeon X1000 and earlier — basically anything from before 2007. We've put an asterisk (*) next to the GPU names for those cards, and they comprise the latter part of the table. Comparing pre-2007 GPUs against each other should be relatively meaningful, but trying to compare those older GPUs against newer GPUs gets a bit convoluted.

These results are, at best, merely theoretical and we don't have any recent benchmarks for most of the GPUs. As one recent example, AMD's RX 7900 GRE ranks above the RTX 4070 Ti Super, even though Nvidia's card nearly matches the RX 7900 XT in rasterization performance and easily beats even the 7900 XTX in ray tracing performance. Take the following with a healthy dose of skepticism and a liberal sprinkling of salt, in other words, but it does contain a list of just about every major desktop GPU from the past 25 years.