Adrenalin 23.2.1: The real RDNA3 GA

AMD released Adrenalin 23.2.1, ending two months of wait with no new drivers for RDNA2 or older GPUs and only urgent fixes or betas for RDNA3. The new driver reunifies all supported GPUs and its Release Notes have a long list of games with improvements and fixed issues.

The question in every RDNA3 user’s mind: Is this the Big One? The new driver that’s fully optimized for the new architecture, that AMD failed to deliver on Dec 13, 2022 with the 7900 series launch?

Checking the release notes again, some items mention support for every Radeon architecture since the RX 400 (like the encoding and streaming improvements); others mention FPS gains specifically for the 6950 XT but that’s probably just a reference point since the same optimization can have impacts of varied magnitude in different GPUs, resolution or other factors. The single exception is for Hogwarts Legacy: +4% FPS at 4K and a bugfix in foliage rendering, both singling out the 7000 series. That’s all great news for owners of every supported Radeon card, another bottle of “fine wine”. But if you’re waiting for optimizations that should unlock unique performance of RDNA3, judging from the Release Notes, this apparently ain’t it.

Benchmarks

I have executed the same tests of my previous stories again, with the launch driver (22.12.1) and the new one (23.2.1). The test platform is fully up-to-date as of 14–02–2023, everything from games to my Asus BIOS.

Some of my results are neutral, with no changes or changes within the margin of error / noise. I will spare you from the noise and only comment results that have significant change for the new driver.

3DMark Mesh Shader
In the 6800 XT GPU that I had before and featured in a previous story, the Mesh Shader feature delivered a massive improvement of +1,531%. The problem was awful performance of this test with Mesh Shaders Off: 35fps. When I first tested the 7900 XTX this disadvantage shrank to +34% — mostly because the test ran much faster with Mesh Shaders Off. But also in part because, with Mesh Shaders On, performance regressed from 559fps in the 6800 XT to 228fps in the 7900 XTX: -59% compared to the older GPU.

The new driver fixes that, catapulting the Mesh Shaders On score to 568fps or +157% over Off. The new score also narrowly beats the older GPU, not an impressive gen-over-gen improvement but at least the big regression observed in the launch driver is gone.

There’s a new, smaller regression: with Mesh Shaders Off, the new driver ‘s score for the 7900 XTX goes down by 17%. Still much better than RDNA2, but it’s a number that goes down from the launch driver. Apparently, there’s still room for improvements in this area. In any case, this feature of DX12 Ultimate is almost completely ignored by games so far, although it might have some adoption for non-game apps such as CAD.

3DMark Gaming Tests

The tests in is group represent game-engine rendering techniques that usually track real-game benchmarks extremely well. See also this article from Igor, relevant beyond memory testing. However, one “limitation” of good, popularsynthetic tests is that these are probably the first things new hardware will be tested and validated for. If it’s true the 7900 cards suffered with poor drivers I’d expect that to impact some games, even many games. But if you launch a GPU that crawls on TimeSpy because of a poor shader compiler or any other driver deficiency that could be fixed — that would be next-level botched. So I expected small impact at best in this set of tests.

Most of 3DMark Gaming tests show modest improvements, around +2%. The one result that exceeds my expectation is SpeedWay, which is nice since that represents advanced ray tracing rendering including global illumination — as in, the games that make RT really worth the cost.

FurMark, Superposition, and Blender
This grouping of tests seems weird: a stress test, a software-RT real time rendering engine, and an offline renderer. They are put together because of common APIs and common results.

FurMark gains +4.58% on the new driver, while Blender improves +3.17%. Both programs are written with OpenGL. Superposition gets +2.26% only with OpenGL — its DX11 variant ties with the old driver. It seems that the new driver has some speedup that’s specific to the OpenGL API.

Real-Game Tests

Despite some neutral results, the full set of scores is a good overview for this section. I included the Quality-level upscaling for games that support FSR 2 or an alternative that works well on Radeon GPUs.

My list of game tests includes good speedups: up to +4% for Quake II RTX, +7% for Guardians of the Galaxy, +5% for Batman Arkham Knight, +3% for Wolfenstein Youngblood. Some of my games are also listed by the driver’s release notes: +6% for GotG, +7% for Quake II RTX, +4% for Hitman 3, +6% for SotR. The discrepancies may be a factor of testing conditions: AMD only includes numbers for the 6950 XT, and only at 4K where driver optimizations often have a higher impact due to that GPU’s 256-bit bus.

I also have significant gains on games like Batman and Youngblood, both old enough that no GPU vendor cares to optimize them anymore or even mention wins in the release notes. That’s evidence that the speedups come from core driver changes, not just game-specific tweaks.

Summary

This is not yet the mythical new driver that would boost RDNA3 GPUs by another +20% in all games and meet the pre-launch expectations. However, an extra +5% (in average) is a good start. My estimation is that this release mostly brings RDNA3 support to GA quality.

As the release notes make clear, most improvements in this driver aren’t even RDNA3-specific. It’s clear the Radeon team spent the last two months putting out the fire of the poor launch driver, I don’t think there’s much work yet in RDNA3-specific optimizations for any games. There’s also significant progress in other features, from encoding and streaming to the MLIR/IREE support for ML applications. But in gaming performance, most improvements in this release are optimizations or driver architecture changes that apply to many supported GPUs, not unique to RDNA3.

Some users reported more impressive gains in ray tracing for RDNA2 GPUs, while others found that the driver might mess with CPU overclocking (a bug that couldn’t reproduce; I use RyzenMaster for CPU OC and the new Adrenalin driver didn’t interfere with that). I look forward to professional reviewers testing this on a wider range of GPUs, games, and settings; also power usage and overclocking. Meanwhile, if AMD just released a new driver that significantly increases the performance of over two year old RDNA2 GPUs across the board, here’s hoping this release is just the first step for RDNA3. As we get closer to the launch of higher-volume, lower-cost entry/midrange 7000 SKUs, driver quality will be even more critical.

Update: One day after the new driver, AMD also released FSR 2.2 — another part of the half-baked Dec 13 launch, when they “released” this only in beta form for two games and no update to the public sources or documentation. This looks like a big step for AMD’s upscaling tech, and the sources hint at improved GPU-specific optimizations (Wave32 vs. Wave64 code paths).