AMD Radeon RX 7600 Review: Incremental Upgrades

The AMD Radeon RX 7600 brings the latest-generation RDNA 3 architecture to the sub-$300 realm. It may even qualify as a "budget" graphics card, depending on how you want to stretch that definition. Originally planned for a $299 launch price, AMD dropped the official RX 7600 MSRP to $269 over the past weekend, likely thanks to Nvidia's RTX 4060 $299 announcement. Numerous leaks over the past few weeks have already hinted at what we'll see today, but let's get to it.

Will the RX 7600 rank among the best graphics cards? We'll give it a resounding maybe. Ultimately, it will come down to street prices over the coming months. Right now, plenty of existing RX 6650 XT, RX 6600 XT, and RX 6600 cards largely fill the same role, but with lower pricing. Those originally had significantly higher MSRPs, thanks to the GPU shortages and mining that was going on in 2021–2022, but now you can pick up one of the XT models for under $250, and the vanilla RX 6600 starts at just $199.

What does the new RDNA 3 architecture bring to the party? Reworked compute units (CUs) with a big boost in theoretical number crunching prowess, improved FP16 (AI) functionality, and better silicon utilization, according to AMD. In addition to our normal gaming, power, and thermal tests, we also ran the card through some stable diffusion tests. Here's the overview of the specifications.

On paper, there appear to be some major upgrades. For example, the dual compute units have twice as many ALUs (arithmetic logic units) as before, so the RX 7600 has a massive increase in FP32 teraflops: 21.5 versus just 10.8 on the RX 6650 XT, or basically double the compute performance. FP16 throughput is also doubled.

That sounds good in theory, but in practice, the real-world gains seem to be far more limited. For example, the RX 7900 XT has 51.6 teraflops of FP32 while the RX 6950 XT only has 23.7, so a 118% increase in compute. In our benchmarks, particularly for games, the 7900 XT is about 20% faster on average. If the same pattern holds with Navi 33 (spoiler: it mostly does), gaming performance won't be significantly faster.

Elsewhere, you don't get any extra VRAM, and both the RX 7600 and RX 6650 XT use 18 Gbps GDDR6. Even the process node hasn't been radically altered, as TSMC's N6 is a refined variant of the N7 node used for Navi 23. Density did improve, though, from around 47 million transistors per square millimeter (Mtrans/mm^2) to 65 Mtrans/mm^2. However, that's still a far cry from the ~122 Mtrans/mm^2 that Nvidia is getting on its Ada Lovelace RTX 40-series GPUs, or the 152 Mtrans/mm^2 for the Navi 31 (RX 7900-series) GCD.

Besides doubling the maximum FP32/FP16 throughput from the shader cores, one of the biggest architectural changes with AMD's RDNA 3 architecture is the move to GPU chiplets. Except Navi 33 doesn't do that, sticking with a traditional monolithic die approach. That's not really surprising, as GPU chiplets are primarily about cost savings, and adding in extra interconnects between a GCD and multiple MCDs likely only makes sense with wider memory interfaces and stacked L3 caches.

Image 1 of 4

AMD and Nvidia have similar marketing surrounding their new 8GB graphics cards. They're good for 1080p, and lots of people use that resolution (or lower). Where AMD and Nvidia diverge is in the particulars of some of the hardware, as well as memory configurations and pricing.

Nvidia has had tensor cores in its RTX GPUs since 2018, and these can help accelerate deep learning and AI workloads — they're part of the special sauce behind DLSS, for example. Nvidia has also put more effort into making fast ray tracing hardware.

In contrast, AMD has generally focused more on delivering improved rasterization performance, at least for comparable price points. There's also less proprietary hardware and APIs. AMD does have "AI Accelerators" in RDNA 3, which share the FP16 compute functionality of the shaders. They're not quite the same as the full-blown tensor core implementations found in AMD's CDNA/CDNA 2 architectures for data centers, but they do provide some optimizations in throughput compared to accessing the same hardware via shaders. (We'll see some of this in our Stable Diffusion testing later.)

Depending on what product tier you're looking at, AMD also offers potentially more memory. However, that's not the case with the RX 7600, as it's basically near the bottom of the RDNA 3 stack and comes with the same 128-bit interface and 8GB as Nvidia's RTX 4060 Ti and the upcoming RTX 4060. Of course, the RX 7600 also costs less than either of those. The mainstream and high-end AMD RDNA 3 solutions (for desktops) haven't been officially announced yet, but it looks like AMD will probably offer a 192-bit interface and 12GB of VRAM with the 7700-class parts, and then a 256-bit interface and 16GB of VRAM with 7800-class parts. We'll have to wait and see how those stack up in performance against Nvidia's offerings.

Image 1 of 28

In case you're interested, here's the full slide deck from AMD's RX 7600 presentation for reviewers and analysts. There are some performance charts and various talking points, like the fact that the GTX 1060 is still one of the most popular GPUs, even though it's now seven years old.

The thing is, people using a seven-year-old GPU probably aren't even playing the latest games. Some of them might be using laptops as well. If someone hasn't upgraded in seven years, it's doubtful that the launch of a new $269 graphics card will suddenly spur tons of people to upgrade. But if you're in the market, it's good to have options, and that's exactly what AMD and Nvidia are doing with the RX 7600 and (upcoming) RTX 4060.

• MORE: Best Graphics Cards
• MORE: GPU Benchmarks and Hierarchy
• MORE: All Graphics Content