GeForce RTX 3090 Founders Edition: An Absolute Unit
The Nvidia GeForce RTX 3090 Founders Edition now claims the top spot on our GPU benchmarks hierarchy, though where it lands on the best graphics cards is a more difficult question to answer. While the GeForce RTX 3080 kept with the existing pricing structure of the RTX 20-series, replacing the RTX 2080 Super at the $699 price point, the RTX 3090 sees the RTX 2080 Ti's $1200 launch price and raises it another $300. There's an alternative viewpoint, however: The GeForce RTX 3090 is also a replacement for the $2,500 Titan RTX, in which case it's a faster card that costs $1,000 less. Either way, you're going to need some deep pockets if you want to own Nvidia's new halo card.
We already have the Ampere architecture deep dive that details what makes the GeForce RTX 3090 tick. Start there if you want more information on how the 3090 and GA102 deliver new levels of performance. We'll cover the highlights here, but the simple answer is that Nvidia has gone with a smaller process node, more cores, faster memory … plus more power and a higher price. At 350W TGP (Total Graphics Power), this is by far the most power hungry consumer GPU Nvidia has ever released (not counting the dual-GPU cards). It's also the most expensive GeForce branded GPU ever (unless you count the dual-GPU GeForce GTX Titan Z, which I don't).
We have a few additional tests we're still trying to run before we render our final verdict (looking at you 8K, or at least 4K with DSR emulating 8K), so we won't have a final verdict just yet. We also have a few AIB (add-in board) partner cards that we'll be looking at in the coming days, once we've had some time to run all the tests and gather some data.
One thing we can't fully predict is availability, but our hunch is that it's not going to be great. It's actually more than a hunch, now, as Nvidia pre-emptively apologized yesterday for the limited availability of RTX 3090 cards (while at the same time confirming performance expectations). Ugh. Considering the RTX 3080 cards have been selling out as fast as inventory arrives, even though the RTX 3090 costs over twice as much … well, some people are paying nearly $1,500 for RTX 3080 cards. It's probably not going to be pleasant if you have your heart set on a 3090 and didn't put in a pre-order. Just give it some time, and things should eventually get sorted out. Cue mom: "Patience is a virtue!" Thanks, ma, but I'm still working on it.
The short summary of the GeForce RTX 3090 is much as you'd expect. It's theoretically about 20% faster than the RTX 3080 based on specs alone, with over twice the memory. The extra VRAM doesn't really matter much for most games unless you're running at 8K (maybe 5K), but it can prove useful for some professional workloads. The RTX 3090 also requires a beefy CPU to get the most out of the card (we'll be running some additional tests in the coming days to show that, but you can get some idea of what to expect from our RTX 3080 CPU scaling article). This is very much a GPU designed for 4K ultra gaming, and at those settings, it's 12% faster than the 3080 on average in our existing test suite, and 14% faster in our bonus test suite. Drop down to 1440p ultra, and the 3090 is only 8% faster than the 3080. If you're running at 1080p? Don't bother, seriously.
Now, let's dig into the specs and look at what's changed relative to other Nvidia GPUs.
Nvidia GPU Specifications
|Graphics Card||RTX 3090 FE||RTX 3080 FE||Titan RTX||RTX 2080 Ti FE|
|Process Technology||Samsung 8N||Samsung 8N||TSMC 12FFN||TSMC 12FFN|
|Die size (mm^2)||628.4||628.4||754||754|
|GPU Cores (FP32)||10496||8704||4608||4352|
|GPU Cores (INT32)||5248||4352||4608||4352|
|Base Clock (MHz)||1395||1440||1350||1515|
|Boost Clock (MHz)||1695||1710||1770||1635|
|VRAM Speed (Gbps)||19.5||19||14||14|
|VRAM Bus Width||384||320||384||352|
|Tensor TFLOPS FP16 (Sparsity)||142 (285)||119 (238)||130||114|
|Launch Date||September 2020||September 2020||December 2018||September 2018|
GeForce RTX 3090 Architecture: Nearly a Full GA102
Again, our Ampere architecture goes into more detail on the various aspects of the new 30-series GPUs. There are a lot of changes relative to the previous Turing architecture, but here's the highlight reel.
First, GA102 uses Samsung's 8N process technology, which means more transistors in a smaller area than TU102. Looking strictly at overall transistor density, GA102 packs in 45 million transistors per square millimeter, whereas TU102 density is 'only' 24.7 million transistors per square millimeter. That's good, but clearly not as good as TSMC's N7 node: The larger GA100 chip used in Nvidia's A100 contains 65.4 million transistors per square millimeter. Also, AMD's Navi 10 has 41 million transistors per square mm, so Nvidia has at least matched that level of density — comparing across architectures is definitely looking at apples and oranges, though.
Moving on, the GeForce RTX 3090 uses a nearly complete GA102 chip. Of the 84 potential SMs (streaming multiprocessors), only two are disabled. That suggests either the yields are very good … or Nvidia isn't planning to sell nearly as many 3090 chips as 3080 chips. We suspect the second option is closer to the truth, and actual yields are a closely guarded secret these days. The 3090 also has seven GPCs (graphics processing clusters), and the ROPs (render outputs) are now part of the GPC instead of the memory controller, giving the 3090 112 ROPS.
Perhaps a bigger change is that the GeForce RTX 3090 comes equipped with 24GB of GDDR6X memory, this time clocked at 19.5 Gbps (compared to 19 Gbps on the 3080). This is accomplished by enabling the final two 32-bit memory controllers on GA102, and then running 24 chips in half-width 16-bit interface mode. That also means the GDDR6X chips are located on both sides of the PCB (printed circuit board), whereas the RTX 3080 only has memory on the same side as the GPU. There's certainly a question of how much this affects GDDR6X memory temperatures, particularly when half the VRAM isn't actively cooled. However, at present, we don't have a way to measure the GDDR6X chip temperatures.
We've discussed the GPU CUDA core changes quite a bit in previous Ampere articles, because it's one of the most significant differences between Ampere and Turing. Turing included 64 FP32-capable CUDA cores per SM, with an additional 64 INT32-capable CUDA cores (and two FP64 cores for compatibility purposes). The FP32 and INT32 cores were separate datapaths and could be used concurrently. The FP32 cores could also run FP16 math at double the performance — 'fast math' mode.
For Ampere, the FP32 cores lose support for FP16 fast math, which gets shifted over to the Tensor cores — it's still twice the FP16 performance of the dedicated FP32 cores, but only equal FP16 performance overall. Meanwhile, the INT32 cores gain support for FP32 calculations, which means total FP32 performance per SM has doubled. At the same time, INT32 performance per SM has stayed the same. This has some interesting ramifications for overall performance, but basically, about 30-35% of the CUDA core workload in games is INT32 (for address pointer lookups, texture calculations, and other similar work). That means a good chunk of the second datapath will be busy with INT32, so the real-world performance boost will often be less than the raw FP32 TFLOPS number would suggest.
Elsewhere, the tensor cores are another substantial upgrade relative to Turing, with twice the throughput per core. Except there are half as many tensor cores per SM. The 3rd generation tensor cores do add support for fine-grained sparsity, which can double throughput again relative to Turing when used. Also worth noting is that the 3rd gen tensor cores add support for INT8 and INT4 data types at 2x and 4x the base FP16 throughput.
The video ports on GeForce RTX 3090 FE are the same as on RTX 3080: one HDMI 2.1 port and three DisplayPort 1.4a outputs. There's no VirtualLink port this round, as VirtualLink is basically dead. All four outputs are capable of 8K60 using DSC (Display Stream Compression), a "visually lossless" technique that's actually not really visually lossless (though you probably won't notice at 8K). We've seen other cards with five display outputs, many opting for dual HDMI 2.1, so keep that in mind if you're planning to use a multi-monitor setup.
Other changes (and we're glossing over things here as we've already covered this in the RTX 3080 review, as well as the Ampere architecture piece) include RT cores that are up to twice as fast — or even more in certain workloads (like motion blur). The L1 cache/shared memory capacity and bandwidth have been increased. The L2 cache is also larger than before, and the L1 cache can be configured as varying amounts of L1 vs. shared memory, depending on the needs of the application. Register file size is also increased, and GA102 can do concurrent RT + graphics + DLSS (previously, using the RT cores would stop the CUDA cores).
The raster operators (ROPS) have been moved out of the memory controllers and into the GPCs. Each GPC has two ROP partitions of eight ROP units each. This provides more flexibility in performance, giving the GA102 and RTX 3090 a total of 112 ROPS. Each GPC also includes six TPCs (Texture Processing Clusters) with eight TMUs (Texture Mapping Units) and a polymorph engine, though these come in pairs and Nvidia enables 41 TPCs for the 3090.
Finally, as discussed before, the RTX 3090’s memory controller has a new feature called EDR: Error Detection and Replay. When the memory detects a failed transmission, rather than crashing or corrupting data, it simply tries again. It will attempt to do this until successful, though it's still possible to cause a crash with memory overclocking. EDR allows the potential for higher memory clocks that may not actually perform better as you approach the limits of the memory. That's because EDR ends up reducing total memory throughput when failed transmissions occur.
GeForce RTX 3090 Founders Edition: Design, Cooling, Aesthetics
Nvidia made significant changes to its Founders Edition cards for the RTX 30-series. The new designs still include two axial fans, but Nvidia heavily redesigned the PCB and shortened it so that the 'back' of the card (away from the video ports) consists of just a fan, heatpipes and radiator fins, and the usual graphics card shroud. Nvidia says the new design delivers substantial improvements in cooling efficiency, while at the same time lowering noise levels. The GeForce RTX 3090 FE takes the overall industrial design and aesthetics of the 3080 FE, then presses the 'Get Pumped' button.
The GeForce RTX 3090 Founders Edition is massive, and we mean that quite literally. It tips the scales at 2189g, or 4.82 pounds. That may not be the heaviest single-GPU card ever created, but with a triple-slot cooler, it's definitely the biggest Nvidia-created card. The 3090 FE also measures 313mm (12.3 inches) in length, about 3cm longer than the 3080 FE, and it's 138mm (5.43 inches) tall. That's about an inch more in each dimension than its little sibling, and it's very solidly built — future archeologists will dig these up and probably think we used them to club each other.
As usual, aesthetics are a highly subjective matter. It looks a bit drab in some ways, but there's no flimsy metal shroud that feels like it might get snagged on something and bent. Also, we were mistaken when we said the Founders Edition cards didn't have RGB lighting. The white GeForce RTX logo and lit up 'X' are RGB LEDs. Nvidia just hasn't provided the software to users yet. If you want to go back to the earlier Nvidia-green for the logo, you can do that, or you can try to merge matter and antimatter by going with Radeon red. Or at least, you'll be able to do that once Nvidia releases its lighting utility.
Nvidia provided the images below showing the various parts of RTX 3090 Founders Edition. The PCB is the same as the RTX 3080, only now with all memory slots populated — on both sides of the board. (Nvidia didn't provide a photo of the back of the PCB, unfortunately.) The cooler may take cues from the 3080 FE design, but it's far larger. The PCB only extends about halfway down the length of the card, with the entire back portion dedicated to helping dissipate 350W of power.
The fans on the RTX 3090 are also 'super-sized' relative to the 3080. The 3080 fans are custom 85mm models, though the two fans are different — the back fan is designed to pull air through the radiator and spins counterclockwise, while the front fan (closer to the video outputs) is slightly thinner and also spins counterclockwise, pushing air out the back of the card. The 3090’s fans are 110mm, again with a thicker fan at the back, and you can see how the fins are angled to help pull air through the radiator.
The integrated rims on the fans help improve static pressure as well, which is something we've seen on some other RTX 30-series cards. Clearly, Nvidia and its AIB partners are pulling out all the stops to ensure the new Ampere GPUs stay cool and quiet. From what we've seen so far, they're succeeding, though the factory overclocks on some models are certainly pushing the GPUs to the limit. There have been a few reports of instability with AIB cards, though nothing concrete yet. We didn't encounter any issues with the Founders Edition at stock clocks, though, so if you'd rather play it safe and skip the bling, the 3090 FE might be your best bet.
GeForce RTX 3090: What About Overclocking?
We had less time with the RTX 3090 FE than the RTX 3080 FE, and we're adding some professional workloads to the mix because this is that sort of card. As such, we haven't had time to investigate overclocking fully yet. We'll be updating this section with more details, but let's just make this clear: This is a $1,500 GPU, already pushing the limits of what most people would deem sensible in terms of power and performance. The stock clocks are slightly lower than the 3080 for a reason, and typical boost clocks still hit 1800-1850 MHz in our testing. If you want to shoot for a world record, by all means, break out the LN2 and strip the cooler and have at it. Our initial testing, meanwhile, is focused on what you get out of the box.
Based on what we saw with the RTX 3080 FE, you can likely get a stable 50-100 MHz bump in GPU clocks with a bit of tuning. That will also mean more power, of course, though the cooler and fans seem to be up to the task. For the GDDR6X, we managed a seemingly stable 20.5 Gbps, but we don't currently have the tools to read the individual memory module temperatures. There's concern from some corners of the web suggesting the chips get hot, though our experience so far suggests overclocks to 20.5 Gbps are viable for benchmarks. Still, don't go too crazy — no one wants to fry a $1,500 GPU.
Sorry if you wanted more for the overclocking section; the calendar is a harsh mistress and just didn't provide enough time to run more tests. Again, we'll be updating this review with more information in the coming days.
GeForce RTX 3090: Test Setup
We're using our standard GPU test bed for this review, which consists of a Core i9-9900K CPU, Z390 motherboard, and 32GB of DDR4-3200 memory. We do want to run some additional tests on other CPUs and platforms, similar to what we did with the RTX 3080 FE CPU scaling tests, but that takes about one solid day of testing per CPU. We'll post a follow-up article once we're able to dig into things more.
We're also focusing on 4K ultra performance, with 1440p ultra being a secondary consideration. We've got nine games in our existing test suite, plus 15 additional gaming benchmarks (one more than we did with the 3080). None of the existing gaming tests used ray tracing or DLSS, while the bonus games and benchmarks focus more heavily on that aspect of the cards.
Our comparison GPUs for this review are limited to Nvidia's top-tier cards from the previous generation, and we've included the Titan Xp, GTX 1080 Ti, and the GTX 1080 FE for good measure. The reasoning is simple. First, AMD doesn't currently have anything that can compete with the RTX 2080 Ti; never mind the 3080 and 3090. The Pascal GPUs are present (but not in the bonus games) to give some historical perspective on Nvidia's Titan vs. x80 Ti vs. x80 series. We'll talk about that more in a bit.
We also tested with Microsoft GPU hardware scheduling enabled this time — it didn't massively affect performance one way or the other in our 3080 testing, but we'll make the switch going forward. We've used the results from the RTX 3080 FE with hardware scheduling, but the other GPUs were tested previously without HW scheduling. Again, at 4K ultra and 1440p ultra, the difference between the two settings is generally negligible, though individual games may show up to a 4% delta in either direction. Overall, the 3080 FE was 1.2% faster with HW scheduling at 1440p ultra, and 1.8% faster at 4K ultra.
GeForce RTX 3090: Built for 4K and Beyond
While we've run our full test suite on the RTX 3090 FE, it's very much a card that needs a higher resolution display to reach its potential. At 1080p, without ray tracing effects, the 3090 is only a few percent faster than the RTX 3080. Even at 1440p, the gains are usually less than 10%. Which isn't to say that the 3090 suddenly rockets ahead at 4K ultra, as it's still only 10-15% faster than a 3080 by Nvidia's own admission. But now we can confirm Nvidia's results.
Let's start with the existing (older / lighter) games before moving into the bonus round. For these games, we're limiting our charts to the 4K ultra and 1440p ultra results. What about 5K or 8K gaming? We're looking at ways to test that, but presently lack the necessary display (and time) to make it work. Again, stay tuned for additional coverage.
Right from the start, this is clearly a case of diminishing returns as far as gaming is concerned. Across our older test suite of nine games, even running at 4K ultra, the GeForce RTX 3090 FE is only 10.5% faster than the GeForce RTX 3080 FE. Drop to 1440p ultra, and the performance advantage shrinks to a mere 8%. Part of the problem is the GPUs are so fast that they're hitting CPU limitations, even at 4K ultra. The other difficulty is that at least several of the games are a bit older and lighter.
We'll see in the bonus games that cranking up graphics settings — in particular with games that support ray tracing effects — increases the gap. However, even at 4K with DLSS, the 3090 FE doesn't create a gap that would fully justify its price premium. That's generally been the case with Nvidia's top GPUs, going back to the first Titan cards and the GTX 700-series, though. So let's look at history.
The RTX 3090 is 'only' 10.5% faster than the RTX 3080, at slightly more than double the price. That's rough, but the old Titan cards were worse. Titan RTX was only 4.5% faster than the RTX 2080 Ti, at more than double the price, and cost $1,000 more than the RTX 3090. The Titan Xp did a bit better, with 6.9% higher 4K ultra performance than the 1080 Ti, and it 'only' cost 71% more.
But that's the catch. The Titan Xp was the same price as the 2080 Ti; the RTX 3090 isn't a Titan, and it still costs $1,500. If we look at the 2080 Ti compared to the 2080 FE, it was 29% faster for 70% more money. Meanwhile, the 1080 Ti was 32% faster than the 1080 and took over the 1080 Founders Edition price point. Or one step further back, the 980 Ti was 33% faster than the GTX 980, for $100 more (but came nine months after the 980 launch).
Historically, then, the RTX 3090 is one of the worst values ever for a GeForce card … and it's one of the better values for a Titan-class card. It's still a massive price premium, and it's not even that beneficial for gaming — twice the price for 10-15% more performance. That's a bigger jump in performance than any previous Titan card has offered, though. We'll dig into performance in professional applications once we're through with the games.
Borderlands 3 is one of the more demanding games, and the GeForce RTX 3090 gets a correspondingly higher benefit compared to the 3080. At 4K ultra it's 14.5% faster than the 3080. It's also 14.6% faster than the 3080 at 4K medium, and at 1440p ultra, it's still 12.5% faster than a 3080. Compared to the Titan RTX, which it sort of replaces at a lower price, the 3090 delivers 55% more performance. It's also just over 60% faster than the RTX 2080 Ti. Overall, Borderlands 3 is the best showing for the 3090 in our current test suite, even though it's an AMD-promoted game.
The Division 2 is another AMD-promoted game, but this time the 3090's performance is basically right in line with the overall average. It's 11% faster than a 3080, but only at 4K ultra — 1440p ultra or 4K medium drops to single-digit percentage point improvements. If you're coming from an RTX 2080 Ti, the RTX 3090 is nearly 50% faster at 4K ultra, and around 40-45% faster at 1440p ultra and 4K medium. It's also 41% faster than the Titan RTX at 4K ultra, and a 30-40% improvement at lower resolutions and settings (besides 1080p medium).
Far Cry 5 is one of the lighter games in our test suite, and also one of the older games — Far Cry 6 is coming next February, and Far Cry New Dawn came out in early 2019 (but with similar performance to its predecessor). The GeForce RTX 3090 still gets 11% higher performance than the 3080 at 4K ultra, but that's a pretty tough sell, and the lead drops to just 3-4% at 1440p and 1080p, if you're wondering.
Final Fantasy XIV is also ready for retirement from our suite, though it continues to scale pretty well with GPU performance. The RTX 3080 FE breaks 120 fps at 4K ultra, which is a first, and it's nearly 15% faster than the 3080. Then again, even a GTX 1080 Ti has been pulling more than 60 fps in FFXIV since 2017.
Forza Horizon 4 wins the prize for the least demanding game in our test suite, as far as relative performance goes. The RTX 3090 is only 9% faster than the 3080, so unless you have an 8K display or maybe a 5K display, such a GPU is overkill. It's also capable of pulling over 144 fps at 4K ultra, which means anyone with a fancy 4K 144Hz HDR G-Sync Ultimate display could actually put it to good use.
We tested Metro Exodus without ray-tracing or DLSS here, though we'll revisit it below with RTX and DLSS enabled. Given the RTX support, it's no surprise this is an Nvidia-promoted game. The RTX 3090 FE ends up delivering 14% higher performance than the 3080, again one of the better showings for the new Not-Quite-A-Titan.
Red Dead Redemption 2 proved to be one of the better games for GPU performance scaling with the 3080 FE, but apparently we've hit the limit with the 3090 FE. It's only 8-10% faster, depending on the resolution and settings you run. It's also the only Vulkan game where performance didn't end up significantly higher than most of the DX12 games — ray traced or otherwise. We're not quite at maxed-out settings (there are about 27 settings you can tweak, and it all has to be done basically manually for consistency), but whatever. At 4K, you can now comfortably clear 60 fps on the 3090, and performance is 40-45% higher than the previous-gen Titan RTX and RTX 2080 Ti.
Shadow of the Tomb Raider is another RTX-enhanced game that we'll revisit below in our bonus round. Running in traditional rasterization mode at 4K, the 3090 ends up 13% faster than the 3080, so it's another decent showing. Drop below 4K, however, and you're looking at a negligible 6% or less benefit. That's better than SLI in some games, which is now basically dead. RIP, SLI — though Shadow of the Tomb Raider is one of the few games that actually supports it, so, you know: dual RTX 3090s for $3000!
Last is Strange Brigade, a game that uses the Vulkan API and also tends to hit very high framerates. For example, the RTX 2080 Ti already manages 114 fps at 4K ultra, but the 3090 jumps to 181 fps — nearly 60% faster and another use case for 4K 144Hz displays. It's also the biggest lead for the 3090 over the 3080 so far, at 16%. You still get good scaling even at lower settings, as well — over 300 fps at 1440p ultra, and 380 fps at 1080p ultra. Too bad the game isn't a popular esports title.
GeForce RTX 3090: Bonus Round 4K Ultra, Ray Tracing, and DLSS Gaming
As with the GeForce RTX 3080 FE review, we ran additional tests on some other games and graphics benchmarks. This includes all the major (for the US anyway) ray tracing enabled games, several DLSS-enhanced games, and a few other recent releases. Oh, and we added Crysis Remastered because everyone loves a good meme. Can it run Crysis Remastered? Yes! Though not at 4K with maximum quality and 60 fps.
Starting with the overall average again, we took the average fps of all 15 games and graphics tests. We ran some of the tests at multiple settings and included all results in the average. With this expanded set of games, the RTX 3090 FE stretches its lead over the 3080 just a bit to 14.4% — still within that 10-15% range.
The RTX 3090 is also 43% faster than the Titan RTX, 50% faster than the RTX 2080 Ti, and nearly double the performance of the RTX 2080 FE. If you're thinking about upgrading from any of those earlier GPUs, you'll definitely notice the increased performance. But then, you'd notice the jump to the RTX 3080 as well. Let's look at the individual results.
We captured frametimes from 3DMark Port Royal rather than using it's built-in score, since we want minimum fps as well. The RTX 3090 is the first GPU to break 60 fps at stock clocks in Port Royal, and it comes out 15% ahead of the 3080. Generally speaking, it ends up being a pretty decent estimate of overall performance with ray tracing, though we wouldn't want to rely on any single benchmark.
Battlefield V with DirectX Raytracing (DXR) reflections and DLSS enabled was barely above the 60 fps mark when it first came out, using an RTX 2080 Ti. With the 3090, it's now close to hitting 100 fps. This is in a singleplayer mission, however — we know the multiplayer mode can be quite a bit more demanding, but even with ray tracing enabled, you should now be able to get 60 fps or more in most areas.
The Boundary Benchmark from Studio Surgical Scalpels is a forward-looking test at what games might be like in the coming years. It includes ray-traced global illumination, reflections on opaque and translucent surfaces, shadows, and ambient occlusion. It also supports DLSS in performance, balanced, and quality modes, but since we were running at 4K, we just stuck with the performance option (i.e., AI-enhanced upscaling from 1080p). Even the 3090 still can't quite break 60 fps, and minimums sit at just over 30 fps. It's about 15% faster than the 3080, though, which is a recurring theme with the most demanding games.
Bright Memory Infinite is another new benchmark, this time from FYQD-Studio. As with Boundary, it's eventually supposed to become a full-blown game, but that might take a year or two (or more). It includes a bevy of ray tracing options: reflections, refractions, shadows, ambient occlusion, order-independent transparency (OIT), and caustics, plus it supports multiple-bounce rays at higher settings, which makes for more realistic reflections and refractions at the cost of performance. We tested using the high and very high presets, again with DLSS 2.0 in performance mode (1080p upscaled to 4K).
Average performance of the RTX 3090 just squeaks past the 60 fps mark, still with dips into the 40-50 fps range. Performance is also about 15% higher than the 3080 again. So far, 'limiting' our testing to only 4K ultra and below, we've yet to find a game where the 3080's 10GB of VRAM is a problem. That doesn't mean such a game doesn't exist, or won't exist at some point in the coming years, but even the most demanding games and graphics demos of today typically don't require more than 10GB at 4K.
Call of Duty Modern Warfare includes ray tracing support, but only for shadows. It's also a bit distracting (to me, anyway) that the player character doesn't have a body or cast a shadow, making the RT shadows a less-than-stellar implementation. The 3080 managed to clear the 60 fps hurdle at 4K and maximum quality, and the 3090 builds on that lead by 15%. It's fast enough that you could enable ray tracing and still do fine on the multiplayer mode. Or, you know, set everything to minimum quality and play on a 360Hz display and maybe gain an advantage that way?
Like Bright Memory and Boundary, Control pushes more than one major ray tracing effect: reflections, diffuse lighting, transparent reflections, and ambient occlusion in this case. It was the first game to do so, and also one of the first implementations of DLSS 2.0. The RTX 3090 FE can basically manage maxed-out settings at 4K with 1440p DLSS upscaling and still maintain 60 fps (with a few dips below that mark). Or, if you use 1080p as the render resolution and upscale that to 4K, performance jumps to nearly 100 fps — about 13% faster than the 3080. It's also over twice the performance of the RTX 2080, something the 3080 didn't quite achieve.
Yeah, we hear you: "But can it run Crysis?" The original shouldn't be a problem, but Crysis Remastered is another beast entirely. We tweeted about this the other day, though, and the ray tracing in the game seems to be … well, it's hard to spot the enhancements in a lot of areas. But the performance hit is still very much present. At 4K using the maximum "Can It Run Crysis?" settings, the 3090 plugs along at just 41 fps. That's still 20% faster than the RTX 3080 — our best result overall — but not really a smooth gaming experience. Supposedly a DLSS patch is coming, which might make 4K playable on the 3090, but we'll have to wait and see.
If you turn off the game's ray tracing setting — which interestingly provides a weird hodgepodge of software ray tracing and hardware-accelerated calculations, only it's built on top of DirectX 11 — performance jumps by over 50%. That's still only 65 fps on the 3090, and it will be interesting to see how long it takes before we have GPUs that can fully max out Crysis Remastered and average 60 fps or more at maximum settings. (My bet's on the RTX 5080 or equivalent, in about four years.)
Death Stranding ran quite well at 4K ultra settings even on RTX 20-series cards using DLSS 2.0, so the RTX 3090 unsurprisingly breezes along with a cool 146 fps average. That's 11% faster than the RTX 3080, and this is certainly a game where 8K with DLSS Ultra Performance mode could deliver smooth framerates. If you have an 8K TV and the game gets a DLSS 2.1 patch, at any rate.
Doom Eternal is another game that uses the Vulkan API, and performance is excellent across all the tested GPUs. Even at the maximum ultra nightmare settings, everything we've tested here easily breaks 60 fps. The RTX 3090, though, is getting very close to 200 fps and delivers 15% more performance than the RTX 3080.
Our Microsoft Flight Simulator 2020 performance analysis found this to be one of the most taxing games around, particularly for CPUs. Still, having lots of VRAM and running at 4K ultra mostly pushes the bottleneck back to the GPU. The RTX 3090 delivered 46 fps, which is still 13% faster than the 3080 and gets close to the 52 fps limit of the ultra settings that we ran into at 1080p and 1440p. The game has a DX12 patch supposedly coming, which might boost performance beyond 60 fps at ultra quality, but we'll have to see it to believe it.
We also did a Horizon Zero Dawn performance analysis and found it to be more demanding than Death Stranding, despite using the same engine. Blame the port, or the open world design, or whatever. The 3090 is the first GPU to fully eclipse 60 fps at 4K ultra — meaning even minimum fps stays above that mark. Performance is also 16% higher than the 3080, and minimum fps (99th percentile frametime fps) is 25% faster than the 3080, suggesting this might be a game that's actually making use of more than 10GB — or that it's caching more data into VRAM to smooth out the occasional stutter, more likely.
We're back to Metro Exodus, this time with ultra quality ray-traced global illumination and DLSS 1.0 enabled. At these settings, Metro is still a beast to run, and even the 3090 and 3080 barely break 60 fps. The 3090 is 13% faster, but what we'd really like to see is DLSS 2.0 support and maybe some additional fine-tuning of the ray tracing effects. Sadly, that ship has likely sailed — or if you prefer, the train has left the station — as 4A Games is working on other projects these days.
Our Project CARS 3 performance analysis has results from a bunch more GPUs and CPUs, and we wanted to see how the 3090 stacked up. Surprisingly, this is another big win for 3090 and its 24GB VRAM, with performance that's 20% higher than the 3080 and 99th percentile minimum fps that's 22% faster. The game seemed to be more CPU limited than others we've tested, but at 4K ultra, it still makes good use of the fastest GPUs available.
Shadow of the Tomb Raider was only the third game to add RTX enhancements back in 2019, but they arrived over six months after the game launched — too little, too late for many. Still, at 4K highest settings with ultra ray traced shadows, we've yet to fully surpass the 60 fps mark. Until the RTX 3090, that is. With minimum performance of 64 fps, 4K highest settings with ray tracing and DLSS also falls to the 3090, which bests the 3080 by 14%. Again, it would be nice if some of these early DLSS 1.0 games could get DLSS 2.0 patches, but that seems unlikely at best. It would also be interesting to see some games go back and add more than one RT effect — the devs could chalk it up as practice for the sequel! Dreams are free at least.
Last we have Wolfenstein Youngblood, the fourth Vulkan game and the first to implement VulkanRT effects. With DLSS 2.0 running in quality mode, the 3090 ends up 15% faster than the 3080 and just about makes it past the 144 fps mark. Drop to DLSS performance mode and performance shoots up to 177 fps — again, 15% faster than the 3080 — so any 4K 144Hz displays can finally strut their stuff. We do have to wonder what it is about Vulkan that often leads to markedly better performance. Is it the API, or just the developers putting in more time on optimizations? Either way, three of the four Vulkan games we tested (Red Dead Redemption 2 being the exception) boasted impressive performance figures.
GeForce RTX 3090: Pro-Viz Benchmarks and Compute
Nvidia says the GeForce RTX 3090 isn't necessarily a gaming GPU. Never mind that it's part of the gaming brand by virtue of the GeForce name; Nvidia claims the RTX 3090 really comes into its own when doing professional content creation workloads. We've busted out half a dozen professional content creation applications and run them on our test PC. We've also run SPECviewperf 13 in both 1080p and 4K mode, which means a bunch more charts. The benefits of the RTX 3090 varied quite a bit, and we did have a few oddities in our testing (DaVinci Resolve performance was lower than expected, and we're still trying to track down the problem). Here are the preliminary results.
This test comes via Nvidia and uses the Kick.exe utility to render Sol — the dancing robot from Project Sol, which was part of the RTX 20-series launch. Except, it's not rendered in real time and instead uses Maya with Arnold to generate a higher-quality path tracing result. The Arnold is a bit unusual, in that the RTX 3090 was barely faster than the RTX 3080. We're still investigating things, as something on our test PC may be incorrectly configured or interfering with performance. Even so, performance was 33% faster than the Titan RTX. Also, by way of comparison, a Core i9-9900K took 902 seconds to render the scene.
Next up is Blender 2.90, and we're using both the Blender Benchmark utility (in Optix mode, not CUDA) along with several variations of an Nvidia-provided scene called Droid Race. The interesting bit with Droid Race is that besides a normal rendering mode, there are two options for handling motion blur. The "GPU Nodes" mode supports Ampere's new GPU-accelerated vector motion blur that delivers virtually identical results in less time. You can flip through the charts of the Droid Race tests along with the six Blender Benchmark scenes to see how the different GPUs compare. Overall, the RTX 3090 ends up besting the RTX 3080 by 26%. More importantly, it's over twice the performance of the Titan RTX.
DaVinci Resolve uses GPU acceleration to handle complex video filters. We have a test project, again provided by Nvidia, but our CPU utilization was very low (15-20% most of the time), and performance was lower than the expected numbers. We're still trying to figure out the root cause, but we ended up with a five-way tie in the 'Wedding Heavy Styles' scene (ignore the decimal point stuff — it was to avoid problems with our chart generation), while the 3090 was 19% faster in the 'Bride Selective Color' scene.
For ChaosGroup's Octane rendering application, we have three different scenes, rendered with and without RTX acceleration. The speedup from RTX ray tracing is typically 2-4x, and Octane is used by many rendering farms for 3D animation. Whether in RTX or non-RTX mode, the RTX 3090 is around 20% faster than the 3080. Go back to the Turing generation, however, and the RTX mode shows much greater benefits for Ampere. The 3090 is 40-50% faster than the Titan RTX in non-RTX mode, but 70-100% faster once RTX is enabled. The gap is even wider against the RTX 2080 Ti, with the 3090 leading by 94% on average.
There are two tests here: one using GPU accelerated H.264 encoding and the second rendering a preview of a clip within Premiere Pro. The encoding test ends up being pretty much a five-way tie among the GPUs. The 2080 FE comes in first, oddly enough, followed by the 2080 Super and then on down the line — minor differences in GPU clock speed are the likely factor. The other test shows more of a benefit for GPUs with lots of memory. The RTX 3090 completed the 'in to out' preview render 22% faster than the Titan RTX, but it was 44% faster than the RTX 3080.
Next up, we have two rendering tests for V-Ray, both using GPU acceleration. The RTX 3090 again claims the top spot, beating the 3080 by 24% the Titan RTX by 70%, and more than doubling the performance of the RTX 2080 FE.
Last, we have the industry standard SPECviewperf 13, which we ran in both 1080p and 4K modes. Several of these test applications benefit from some of the features enabled on the Titan RTX (but not on GeForce cards), so there are situations where Titan RTX still ends up faster — and there are other professional tests where the Quadro cards would be even faster still. Where are the Quadro Ampere cards? They haven't been announced yet, but Nvidia is holding its GPU Technology Conference (GTC) from October 5-9, which seems like a good place to announce new Quadro cards.
By the numbers, the RTX 3090 ended up beating the RTX 3080 by anywhere from 4-21% in the SPECviewperf 13 tests. The biggest leads were in catia-06, energy-02, and medical-03, while the smallest lead was in sw-03 at 4K. Overall average performance was 11% faster than the 3080. The Titan RTX really mixes things up, however.
Compared to the Titan RTX, the RTX 3090 is up to 32% faster (in maya-05, medical-02, and showcase-02). It does a bit better at 4K than 1080p in some other tests … but then there's catia-06 and snx-03. The Titan RTX is nearly twice as fast as the 3090 in catia-06, and about 20 times faster in snx-03. Obviously, if either of those software packages are things you use, the 3090 isn't going to be a great fit. Overall, for SPECviewperf, the 3090 ends up 35% slower than the Titan RTX, but it's heavily skewed by the massive gulf in snx-03 performance.
As a potential professional GPU solution, the RTX 3090 FE has a lot of promise. That's especially true for AI research, where the deep learning performance of GA102 at least comes somewhat close to the far more costly Nvidia A100. The large size of the RTX 3090 FE does present some problems, however. While we wouldn't recommend using dual 3090 cards for SLI and gaming, multi-GPU solutions aren't as uncommon with proviz workstations and AI research. The difficulty would be figuring out how to cram more than two RTX 3090 FE cards into most cases. Liquid cooling or other options might be required, and we'll be interested in seeing what Nvidia does with its Ampere Quadro cards. We've already heard of at least one manufacturer who plans on putting up to four RTX 3090 cards in a workstation, which could potentially beat even the A100. Scaling out to hundreds of GPUs is a different story, but for a more modest setup, there's a lot of potential in the RTX 3090.
Is the RTX 3090 Any Good for Cryptocurrency Mining?
One last compute-related test to mention is cryptocurrency mining. Yeah, we know, but it's a possibility. The good news (not for miners) is that even at 105 MH/s, which is what the RTX 3090 manages in Ethash, that would only net about $2.50-$3.00 per day in profit — assuming prices don't crash or spike, naturally. That's already down to about half of what we saw in terms of profitability at the time of the RTX 3080 launch, and with the higher entry price it simply doesn't make much sense to use a 3090 for mining. 500 days or more of 24/7 mining to break even? Yeah, no thanks. Not to mention, it's not at all good for the environment to burn energy just crunching random numbers. Please, just leave the GPUs for the games and imaging professionals. #GrumpyGamer
GeForce RTX 3090: Power, Temperatures, Fan Speed, Clocks, and Noise
The GeForce RTX 3090 comes with the highest TGP we've ever seen: 350W. It also comes with the biggest heatsink I've personally seen, though Zotac's Amp! Extreme line is a close second. Let's see what the actual power, thermals, fan speeds, clocks speeds, and noise levels look like under Metro Exodus at 1440p ultra, and while running FurMark. We use Powenetics in-line power monitoring hardware and software to collect the real power use of the graphics card. You can read more about our approach to GPU power testing, with the main point being our methods are not reliant on drivers or software to gather the power data (though we do use software for the fan speeds, GPU core temperatures, and clock speeds).
How much power does the GeForce RTX 3090 Founders Edition use? Like the RTX 3080 FE, the 350W TGP rating ended up being a bit conservative. We hit 360W averages in both Metro Exodus and FurMark. Here's the full set of data, but it's also worth noting that noise levels on the RTX 3090 FE are basically minimal — the noise floor for our test PC (at a distance of 15cm) is 43 dB, and under load we saw a maximum of 48.9 dB — just a hair lower than the Asus GeForce RTX 3080 TUF Gaming. It's also slightly quieter than the RTX 3080 FE's 49.7 dB. From above the desk, the load vs. idle noise levels aren't really noticeable and run into the limits of our test environment (which most definitely is not an anechoic chamber, sorry).
Temperatures, fan speeds, clock speeds, and noise are all interrelated and affected by power consumption. Of course, a larger heatsink and fans can cool better than a smaller heatsink, and the 3090 puts its size to good use, allowing it to consume more power than the 3080 FE without getting hotter — in fact, it runs a bit cooler.
We'll have an FLIR thermal video later that shows a timelapse 20 minute video of the RTX 3090 FE launching and running Metro Exodus and then exiting back to the desktop, but the exterior of the card tops out at around 55C, which is a far cry from the scorching backplates of the faster 20-series Founders Edition cards.
Overall, the RTX 3090 FE does great when it comes to cooling performance and noise levels. The RTX 3080 FE does pretty well, but we've already seen a couple of AIB cards deliver superior temperatures and noise levels. The RTX 3090 FE basically matches the Asus and MSI 3080 cards and tops out at a not-particularly-worrisome 65C, with the fans spinning at just 1100 RPM. We know other 3090 cards have equally massive triple-slot designs, and some of them may end up beating the 3090 FE, but there's little cause for concern.
The large size of the RTX 3090 FE does present some problems, however. While we wouldn't recommend using dual 3090 cards for SLI and gaming, multi-GPU solutions aren't as uncommon with proviz workstations and AI research. The difficulty would be figuring out how to cram more than two RTX 3090 FE cards into most cases. Liquid cooling or other options would be required, so we'll certainly be interested in seeing what Nvidia does with its Ampere Quadro cards.
GeForce RTX 3090: The New King of the Graphics Card Hill
Let's be clear: the GeForce RTX 3090 is now the fastest GPU around for gaming purposes. It's also mostly overkill for gaming purposes, and at more than twice the price of the RTX 3080, it's very much in the category of GPUs formerly occupied by the Titan brand. If you're the type of gamer who has to have the absolute best, and price isn't an object, this is the new 'best.' For the rest of us, the RTX 3090 might be drool-worthy, but it's arguably of more interest to content creators who can benefit from the added performance and memory.
We didn't specifically test any workloads where a 10GB card simply failed, but it's possible to find them — not so much in games, but in professional apps. We also weren't able to test 8K (or simulated 8K) yet, though some early results show that it's definitely possible to get the 3080 into a state where performance plummets. If you want to play on an 8K TV, the 3090 with its 24GB VRAM will be a better experience than the 3080. How many people fall into that bracket of gamers? Not many, but then again, $300 more than the previous generation RTX 2080 Ti likely isn't going to dissuade those with deep pockets.
Back to the content creation bit, while gaming performance at 4K ultra was typically 10-15% faster with the 3090 than the 3080, and up to 20% faster in a few cases, performance in several professional applications was consistently 20-30% faster — Blender, Octane, and Vray all fall into this group. Considering such applications usually fall into the category of "time is money," the RTX 3090 could very well pay for itself in short order compared to the 3080 for such use cases. And compared to an RTX 2080 Ti or Titan RTX? It's not even close. The RTX 3090 often delivered more than double the rendering performance of the previous generation in Blender, and 50-90% better performance in Octane and Vray.
There are still situations where the Titan RTX wins out, however. Across the SPECviewperf 13 suite, the Titan RTX comes out ahead in overall performance. There are a few cases where the difference is night and day — Siemens NX 8 is the big one, where the Titan was about 20x faster than the RTX 3090, but it was also about twice as fast in CATIA V6. If you're consistently using certain professional applications, of course, you can probably justify moving up to a Quadro card.
The interesting thing is that as fast as the RTX 3090 is, Nvidia could have gone faster — for a higher price, naturally. TSMC's N7 and N7P nodes are in full production now, and Apple has chips coming off the N5 node. Samsung's 8N is more like 10nm++ and is a step down in density and power requirements. But with a line of buyers already soaking up TSMC's more advanced nodes, it can afford to charge more. Samsung was clearly a better value prospect for Nvidia, and at least for the RTX 3080, we end up with an extremely impressive GPU at a not-entirely-unreasonable price. The RTX 3090, meanwhile, claims the top honors for performance for anyone who's willing to pay for it.
Given what we've seen of Ampere so far, we're also excited to see what Nvidia's GeForce RTX 3070 will have to offer. We're still skeptical about it matching the RTX 2080 Ti performance in all games, but it should come close, and at a much more affordable price. It's also a 220W part, so you won't have to worry about a PSU upgrade if you have a relatively decent build.
The wildcard in the best graphics card competition is AMD's Big Navi. Current details say there will be 80 CU models with 16GB of memory, but on a 256-bit bus. The ray tracing performance for RDNA2 is also a massive question mark — Nvidia's Ampere can be up to twice as fast as Turing in some ray tracing tasks, while there are hints from the Xbox Series X and PlayStation 5 that suggest AMD might be taking a far more cautious approach to ray tracing.
Either way, we'll find out what those other cards can do soon enough. The RTX 3090 will probably remain at the top of the GPU benchmarks and performance hierarchy unless AMD can pull a rabbit out of its hat, but RTX 3080 levels of performance aren't out of the question. If AMD can come close to RTX 3080 or maybe even exceed it, at a lower price, that would be very interesting.
For gamers, waiting for the dust to settle — and inventory to build up — is a better plan than shelling out $1,500 for the RTX 3090.