Review: The Radeon 7900 XTX

17 min readDec 28, 2022

In this study I look at the Radeon 7900 XTX’s performance compared to my previous GPU, the Radeon 6800 XT. Given all debate about RDNA 3’s release with possibly immature drivers or other problems, I intend to make this an ongoing project for some time with follow-ups after important releases of the driver, benchmarks, or games in my list.

This is a high-effort amateur product. I am experienced in coding and analyzing benchmarks, but not specifically for GPUs. I’ll have some charts here, and you can see the full data as a Google Sheet.

Disclosure of platform and limitations:

Resolution: 3440x1440, except for tests that run at a fixed resolution.
Hardware: 5900X, 32GB, Asus TUF Gaming X570-PRO WIFI II, Samsung 980 Pro 2TB, NH-D15 cooler, 32GB RAM 3800CL16. ReBAR enabled and other BIOS tweaks. Easy CPU tuning / Ryzen Master’s auto per-core CO.
Software: Windows 11H2. Well-groomed for bloat but I have most virtualization and security features enabled. Major exceptions of memory integrity and disk encryption, both disabled.

Given my testing on a single resolution, the Ultrawide-1440p is a balanced choice to represent the performance of a high-end GPU: 34% harder than regular 1440p but 40% easier than 4K.

Test methodology is simple:

Last of 2 runs to limit cold-run factors like shader compilation. I’m not collecting 1% lows, which requires more effort to control noise.
Using only synthetic benchmarks and built-in game benchmarks. The latter have limitations, but they are easy to run and reproducible.
All games configured to absolute max settings unless noted otherwise.
Every firmware or software, relevant or not to the tests, updated to the latest production release (for this review: by Dec 26, 2022).

My real-game results are limited to games that have built-in benchmarks, which are a minority, still I own a good number of those. I don’t buy games just for testing so the list is biased to my preferences. With that grain of salt, there’s a decent variety including raster and RT, multiple engines and APIs, modern titles, and some old classics.

My RDNA 2→3 upgrade

My primary motivation for this series of benchmarks was justifying my purchase: is the 7900 XTX good enough to run every game I care about without any significant compromise?

In some cases like heavy RT games I expect the answer to be “yes, with some upscaling”. AMD’s FSR 2 is now a near-zero compromise at the Quality setting and 1440p or higher, so I’m also testing that where available.

Through this document, I will often quote “gen-over-gen” performance scores by adjusting my baseline 6800 XT results with +9%, the advantage of the 6900 XT according to the TechPowerUp database. This is only the average of many games. But that number is very close to the theoretical gap of +11% shader units or FP32 TFLOPS; that’s almost a perfect match, real-world performance never scales perfectly with more silicon.

I have the AMD reference cards for both the 6800 XT and 7900 XTX so I’m not benefiting from AIB models with better clocks, TDP, or cooling.

My expectations

TechPowerUp benchmarked the 7900 XTX beating the 6900 XT by +45% raster / +56% RT. This is the gen-over-gen pairing that makes sense to me: both GPUs are the release-day SKU for the largest, uncut die. Most reviews (including AMD’s launch slides) chose to compare the XTX to the 6950 XT, but the latter was a respin of the 6900 XT coming 6 months later with higher clocks and more bandwidth. I expect a similar update for RDNA 3 (7950 XTX?) but we don’t have that yet. Adjusting TPU’s scores for the 6950 XT (+6%), the 7900 XTX is +43% / +53% RT. Even without that, TPU’s results compare unfavorably to AMD’s averages of +54% / +68% RT. This discrepancy caused much debate, but it’s mostly the effect of AMD picking games that scale well to a faster GPU and emphasizing “up to” maximums.

Adjusting for my tests against the 6800 XT, I should expect +50% raster / +61% RT to match TPU’s averages. Ray traced games are actually a majority of my suite of real-game benchmarks, but most of those games are hybrid raster/RT engines so both techniques are well represented and I didn’t bother to run additional tests with RT disabled where that is an option.

Synthetic and Non-Gaming Benchmarks

Synthetic / Non-Gaming — 7900 XTX vs. 6800 XT

3DMark Raster: TimeSpy, FireStrike
TimeSpy Extreme makes +56% over the 6800 XT or +43% gen-over-gen. Meanwhile, the FireStrike Ultra test improves by 54% or 41% gen-over-gen. These are raster tests, using DX12 and DX11 respectively, at 4K resolution.

Notice that the 1440p variants of the same tests (TimeSpy and FireStrike Extreme) show smaller gains for the new GPU. This is as expected; the 7900 XTX’s 384-bit bus is a significant advantage at higher resolutions, fixing a well-known handicap of Radeon’s previous generation at 4K.

3DMark RT: SpeedWay, Port Royal, DirectX RT
In the hybrid Raster+RT that represents all mainstream RT games today the 7900 XTX gains up to 70% or 55% gen-over-gen. That’s a good start with RT.

In the pure-RT DXR benchmark we have +101% over the 6800 XT, +85% gen-over-gen which beats the +80% estimated from the main architectural changes for RT: +50% ray/triangle intersections per CU x +20% CUs.

3DMark API Overhead
Starting with the caveat: this test is unsupported since Jan 2021, so it might have bugs and limitations that are not addressed anymore. It’s unfortunate since there’s no other benchmark that I know with the same focus and covering all of DX11, DX12, and Vulkan.

For both modern/low-level APIs this test shows another regression for the 7900 XTX, scoring -13% draw calls/s for DX12 and -20% for Vulkan against the 6800 XT. I ran this many times, it’s a consistent result.

Intel XeSS
Intel’s XeSS is sort of multiplatform: only the Arc-specific implementation runs its full ML-based reconstruction system; support for any other GPU is limited in both quality and performance. But you’ll find the rare game that supports XeSS and not FSR 2. In the 6800 XT, XeSS improved performance by +32% in the Ultra Quality preset (1.3X scale in each dimension).

The 7900 XTX however only gets +19%; the test gains +65% with XeSS off and +49% with XeSS on. Because the latter number is smaller, the benefit of using XeSS is narrower. The XeSS pass scales less well in the new GPU than the rest of the frame rendering.

3DMark DirectX 12 Ultimate tests
The three tests here are DX12 features that have small adoption on PC games as far as I know, so few people care about these. But my tests again have some interesting outcomes.

Sampler Feedback makes minimal difference in both RDNA 2 and 3. The animation now runs in excess of 800fps on the XTX, it’s so easy to render that this DX12 feature is not necessary to gain performance. For what it’s worth, the 7900 XTX improves +57% over the 6800 XT.

Mesh Shading makes a difference, but this test shows a big regression: without MS the 7900 XTX wins by +388%, but enabling MS it loses by -59%! The feature still gives a boost of +33%, but the older GPU got a ludicrous 1,531% (in part because its MS-off test ran very poorly). Clearly, there was some change in RDNA 3’s geometry engine or drivers that privileged the common-case of MS-off (we should still celebrate a huge 5X win there!), but sacrificed the performance of virtualized geometry.

VRS, finally, has more normal results, and similar behavior in RDNA 2 & 3.

VRMark
VRMark includes three “rooms” or tests. Two of them (Orange and Cyan rooms) are useless as benchmarks since they are framerate-limited, delivering identical scores for any GPU that’s fast enough at rendering a total resolution of only 74% of 1080p.

The Blue room is the good test here, designed for modern VR devices and rendering at 5K / 2560x2880 per eye or 77% more pixels than 4K. It also has no framerate limit. The 7900 XTX scores 185fps, or +50% over the 6800 XT. This translates to +37% gen-over-gen, pretty good for 5K.

FurMark
This little program is very simple and designed primarily as a GPU stress test, so the relevance of its scores to any real-world workload is small. But it’s a rare popular, OpenGL-based test. The 7900 XTX gains +44% over the 6800 XT (1440p preset). Perhaps just as important, a stress test at my full resolution has the 7900 XTX peaking Tjunction at 80ºC, board power 345W, fans at 1,800rpm (all my tests use stock settings including fan curves).

In this study, I didn’t focus on stress and limits but I checked GPU sensors for all benchmarks and some regular playing of my favorite games. The heaviest ones bring Tjunction to 90ºC with the same maximums for power and fan speed. YMMV with custom fan curves, OC, room temperature (I had 24ºC through all testing), and the silicon lottery. Some users of the reference 7900 XTX reported much higher Tjunction, but this seems to be some manufacturing issue such as poorly-applied thermal paste.

Superposition
This interesting benchmark uses a ray tracing engine that’s not GPU-accelerated (in the sense of using RT cores/accelerators; it still renders on the GPU). With that in mind, the 170fps score from the 7900 XTX looks impressive. That’s +60% over the 6800 XT, +38% gen-over-gen.

Blender
The 7900 XTX scores +41% over the 6800 XT or +29% gen-over-gen. That was inconsistent with other reviews so I decided to investigate.

The benchmark uses Blender’s Cycles renderer, not AMD’s RDNA-optimized ProRender, but that’s how everyone runs it. The 7900 XTX looks relatively faster in some reviews because their scores for the 6800 XT or 6900 XT are much lower than mine. Solving this mystery in steps:

BlenderBench’s database gives the 6800 XT a median of 1947 for Blender 3.3.0 (like most reviews) but 2514 for 3.4.0 (within 1% of my score).
Blender 3.4.0 was released a week before the 7900 series, early enough even for reviewers that received RDNA 3 cards a few days before Dec 13.
BlenderBench’s UI makes you pick the Blender engine at every run, downloading it if necessary. The latest version always comes first.
Most professional reviewers keep a database of past runs, reusing them as long as all components involved had no significant change.

Unfortunately, it seems that some reviewers failed on that last step, mixing old data for the 6800 XT or 6900 XT on 3.3.0 with new data for the 7900 XTX on 3.4.0. I tried to re-test my 7900 XTX with Blender 3.3.0, but that doesn’t work: “Did not receive Benchmark JSON Data” error, that’s why you also can’t find any results for that configuration on their site.

Blender 3.4.0’s release notes mention new support for AMD HIP and the RX 7000 series. The results site also shows only +4% over 3.3.0 for NVidia’s 3000 and 4000 series. In terms of performance, 3.4.0 is an AMD-focused update with a +30% speedup for RDNA 2 or better. That’s very good news since AMD’ support for non-gaming use cases in their client GPUs is still evolving to be generous; an RTX 4080 is still 2.5X faster than the 7900 XTX on 3.4.0. Any double-digit improvement is welcome, but still ways to go before Radeon is competitive here.

This investigation confirms a narrow win of the 7900 XTX over AMD’s previous generation. Perhaps the “support for 7000 series” part of 3.4.0 was preliminary and RDNA 3-specific optimization can improve further.

Game Benchmarks

Gaming — 7900 XTX vs. 6800 XT (Native rendering)

Shadow of the Tomb Raider
In this favorite of GPU reviews, the 7900 XTX is +50% over the 6800 XT or +42% gen-over-gen. Good but not impressive. Having said that, the 7900 XTX does 110fps on this “lightly raytraced” game at 3440x1440 max settings with no upscaling, a good start with RT.

SOTR is also my only real-game test for XeSS, where Quality mode nets only +17% in framerate and Performance mode only +40%. Those results are consistent with the synthetic 3DMark XeSS test for the 7900 XTX.

Quake II RTX
This path-traced remaster of a classic hit was Ampere’s Portal RTX, which I could only run well on my 6800 XT with significant upscaling. Now the 7900 XTX renders this game natively at 62fps, +77% better or +62% gen-over-gen. That’s pretty close to the promise of up to +80% over RDNA 2 in ray tracing, confirming the synthetic 3DMark DXR test.

The game’s upscaling is limited to DLSS, FSR 1, and a conventional TAA; no FSR 2 or XeSS. Fortunately, both TAA and FSR 1 are good enough for a simple game like Quake II, even with the improved renderer. The two options have similar performance; I’m using the numbers for TAA here because I don’t think anyone should care about FSR 1 anymore–I count “FSR 1 support” as “no FSR support”. With TAA at 65% scaling (Quality level for DLSS/FSR/XeSS), the 7900 XTX runs this path-traced title at 129fps.

DOOM 3
This 2004 game is an example of a very old title that uses a GPU-accelerated API that’s still supported (OpenGL). Still, it’s a terrible benchmark for modern GPUs because the game is completely single-threaded, but I wanted to test something like this just in case. (The 2012 BFG Edition remaster doesn’t support the built-in timedemos.) The game runs at 377fps in the 6800 XT, and the 7900 XTX improves that only to 408fps, an almost-symbolic +8% which is better than I expected considering the game.

Batman: Arkham Knight
Arkham Knight had groundbreaking rendering for 2015, still holding up incredibly well with an engine that doesn’t trace any rays. Hot tip: you don’t need good shadows if it’s always dark and rainy. My old GPU could run it maxed out at 193fps; the 7900 XTX scores 242fps, a modest +25%. That doesn’t mean much: it’s a DX11 game hitting its limits; even a 4090 doesn’t get it any faster with the same CPU (source: my YouTube digging).

Still, unlike DOOM 3 this game is far from pre-historic, illustrating the difficulty of benchmarking high-end GPUs or even producing averages that represent all games that are still relevant–I only played the whole Arkham trilogy this year, enjoying it as much as any recent AAA release.

Hitman 3
Among the heaviest games on my list, Hitman 3’s Dubai test crawls at 27fps at max settings in the 6800 XT. This happens in part because Hitman 3 was patched with an implementation of RT that got panned for its heavy cost for modest benefit: only shadows and reflections, a level now considered “meh” by ray tracing enthusiasts.

In any case, the 7900 XTX improves +89% at 52fps, +73% gen-over-gen which is an exceptional result. If every game was like Hitman 3, the 7900 XTX would have launched to universal acclaim despite still falling well behind the RTX 4090 on RT.

Red Dead Redemption 2
RDR2 runs well on the 6800 XT, 89fps at max settings: quite good for a prev-gen, 80-class GPU at one of the best-looking games of all time. The 7900 XTX improves that by +36%, 121fps or a measly +25% gen-over-gen. FSR 2 Performance looks similarly disappointing with +14% for Quality and +23% for Performance, indicating a CPU-limited game.

Wolfenstein: Youngblood
This game supports ray tracing but only using early NVidia extensions, it was never updated to a standard API so it still doesn’t work on any other GPU (the single case I know with that problem). I’m testing in raster mode. The 6800 XT was fast enough with 219fps and the 7900 XTX improves that by +56%, reaching 343fps, +43% gen-over-gen.

Guardians of the Galaxy
The 7900 XTX beats the 6800 XT by 70%, running this maxed-out-including-RT game at 78fps which looks great. To be fair the use of ray tracing here is very basic (only reflections), but the game’s visuals are spectacular and very heavy for other reasons; RT doesn’t have a lock on graphics innovation yet.

Dying Light 2
The first game with an advanced, hybrid raster/RT implementation in my list, Dying Light 2 runs on the 7900 XTX +60% better or +47% gen-over-gen. That’s pretty good, still a bit short of that “up to +80%” in RT.

This is also my first real-game test where FSR 2 gives excellent gains, +57% for the Quality preset and +110% for Performance. With FSR/Quality this game scores 82fps at maxed-out settings for my 3440x1440 test resolution. Captured data shows 100% GPU utilization, as one should expect from a benchmark when the game is not CPU- or engine-limited.

Callisto Protocol
The single AMD-sponsored game in my list, Callisto should have been the star of these reviews–RDNA 3’s RE: Village, showing that Radeon can also push the envelope on ray tracing. Unfortunately, the game shipped with many performance issues. Some problems are fixed already in the version I tested (shader compilation), but more is needed.

The 7900 XTX runs the game at 72fps, +38% over the 6800 XT. That looks great for a game that delivers on the promise of next-generation visuals, however, that performance is for the regular 1440p since the benchmark only runs at the 4:3 aspect ratio; in my actual gameplay at 3440x1440 the framerate is generally above 60fps but it can dip as low as 45fps.

Upscaling should fix that, right? For the 6800 XT, FSR 2 Performance brings the score from 52fps to 71fps, a modest +35% for that level of upscaling. But on the 7900 XTX I get zero lift from FSR 2 even at the Performance preset, in the benchmark or gameplay. In both GPUs and in all configurations, this game hits a brick wall in my system at 72fps. And this is not a review of the game, but the non-tunable motion blur is also a major annoyance. And I experienced two crashes.

Performance capture data shows how the game performs poorly, with low utilization of both the CPU and GPU. (Dark background because of the benchmark’s 4:3 aspect ratio, leaving the left and right sides black.)

More RT

I also looked at other games featuring ray tracing but no built-in benchmark. These are informal tests, just playing through hard scenes (complex levels / combat), only on the 7900 XTX. All settings maxed out.

Control: Solid >60fps, often goes into the low-70’s.
Deathloop: Runs above 100fps, but crashes a lot in some areas.
Doom Eternal: Around 200fps, great even for its lightweight RT.
Fortnite: I tested the Chapter 4 update showcasing Unreal Engine 5.1’s Lumen/Nanite, with visuals that often challenge path-traced engines. It runs at a smooth 60fps without upscaling.
Minecraft RTX: Average 45fps for the lowest quality (render distance 8), 35fps for the highest (render distance 24). There’s no upscaling support other than DLSS so this game was absolutely unplayable in the 6800 XT at my resolution. Now it’s tolerable, but this really needs FSR 2.
Spiderman: Remastered does 75fps, Miles Morales (a step up in RT) does 65fps. FSR 2 makes no difference for either in the averages but helps prevent dips. Both games are notoriously CPU-limited with RT.
Portal with RTX: The Low preset is the only hope to play the game: 20fps native, 55fps on TAA-U/Performance. A game at 55fps should be very playable but Portal feels strangely sluggish, it would appear that some frames are being dropped but still counted. Also, image quality is very poor with these settings. There’s no support for FSR 2. There are compatibility problems as well: the Volumetric Light option causes comical rendering artifacts in the portal gun.
Resident Evil Village: Runs around 180fps, more impressive than Doom Eternal since this game features way more advanced RT including GI and AO (despite compromises such as very low-resolution reflections). Surprisingly for an AMD-sponsored game, upscaling is stuck at FSR 1. But on RDNA 3 you don’t need any.
DOOM-RT: This path-traced mod stays above 80fps with a rendering resolution of 1080p (slightly better than the Quality setting of DLSS or FSR 2, none of those supported, just simple upscaling). But there are very bad rendering artifacts that I didn’t have on the 6800 XT.

Software

Other reviewers and enthusiasts pointed at a variety of games and benchmarks where the 7900 series underperforms their expectations, speculating about poor drivers, limited use of the new dual-issue shaders, clock/power management, VGPR/cache stalls in RT, etc. I’m often guilty of engaging in these idle discussions but here I will focus on concrete findings. Some evidence that RDNA 3 shipped with subpar drivers:

Frequent crashes in Deathloop. A major title, completely broken for me.
Portal RTX has bad rendering issues. Blaming AMD for problems in an RTX techdemo seems unfair, but the game uses the standard DXR API.
DOOM-RT has bad rendering issues as well. This is an amateur project, a mod to PrBoom where the initial versions didn’t even run on Radeon GPUs; still, the 6800 XT could run this, slowly but correctly.
Two 3DMark tests, API Overhead (DX12 and Vulkan) and Mesh Shading, have significant regressions. My old 6800 XT beating the 7900 XTX by >2X is an extremely abnormal result even in a bad microbenchmark.
I can give a pass to HYPR-RX & FSR 3 promised for H1’2023, but FSR 2.2 should have launched simultaneously with RDNA 3. AMD claims having launched it, but all we got so far is two games running prerelease code. The Github repository is still stuck at 2.1.2 as I write this.
(Updated) AMD’s Radeon 7900 XTX RGB Tool, also tasked with the GPU’s firmware updates, was released only on Jan 5, 2023. The same date that AMD already released the first firmware update (V1.00.06).

Nothing in this list is direct evidence that specific games are held back by problems that can be fixed with a driver update, but it’s hard to not see a rushed release and potential for improvement. I will let the benchmarks answer the question, checking if scores move in a significant way with new drivers or game patches.

Performance Conclusions (for now)

If I discard the tests that turned out very CPU-limited in my system, the 7900 XTX is roughly 60% faster than my old 6800 XT for my real-game benchmarks or 50% faster than RDNA 2 in my gen-over-gen estimations. This looks good because my selection of games is heavily biased toward ray tracing. My tests also show more modest gains with pure rasterization.

These results are largely in agreement with major independent reviews. There’s much talk about RDNA 3 failing to meet expectations, but these come from AMD’s benchmarketing and from counting architectural elements like CUs, ray/triangle intersections, or bandwidth. My own view is that RDNA 3 so far underdelivers in raster but basically meets the promise in RT. The heaviest RT tests, including path-traced benchmarks and games, produce the highest gen-over-gen speedups in my selection.

Is RDNA 3 finally good enough for RT? Independent reviewers place the 7900 XTX between the RTX 3090 and the RTX 4080, with an average that ties with the 3090ti which two months ago was the king of ray tracing. So yes, the 7900 XTX is very good. The major problem for AMD’s release is that NVidia raised the bar of RT by an impressive level with Ada.

Part of my motivation for this work (besides that being my idea of fun) is that the current crop of RDNA 3 benchmarks is very preliminary. We can blame that on the late release in a Q4 packed with new hardware, I guess. From my go-to GPU reviewers, the best was Hardware Unboxed / TechSpot with 16 games, but only 5 ray tracing tests and no attempt to produce an average score for this small selection of RT games. TechPowerUp’s review is the most comprehensive so far with 25 raster games, 8 raytraced games, and a unified score for RT that puts the 7900 XTX only 16% behind the 4080 thanks to games like Far Cry 6, RE: Village, and Watch Dogs: Legion; all of them skipped by reviewers that didn’t forget Control or Cyberpunk 2077.

Even that 16% behind is still significant when comparing the top-of-line RDNA 3 GPU with Ada’s second-tier; the latter is 20% more expensive (MSRP) so price:performance is about the same for RT-loving gamers. The ideal scenario for Radeon to compete strongly in this generation would be gaining another ~5–10% in ray tracing, either via driver improvements or other factors such as AIB cards with good overclocking gains.