AMD architecture performance analysis: where GCN 1.0 collapses

Thus far, results have suggested that tracking performance across AMD generations is challenging because of the vast difference in the overall trend of results depending on whether DX11 or DX12/Vulkan is used. It raises the question: is modern AMD hardware under-utilised in DX11, or is GCN 1.0 really that bad under DX12? I suspect that it may well be a bit of both and while there's plenty of conjecture in the following results, the sense is that testing GCN 1.0 in a like-for-like manner with later AMD architectures can present all manner of issues - and not necessarily down to API support.

Speaking to game developers about PS4 Pro in particular, the impression we get is that the system is somewhat unbalanced - mostly owing to paltry upgrades in memory bandwidth and VRAM allocation. However, geometry processing sees a huge upgrade over the base PlayStation 4, something we see in the later architectural designs from AMD. Perhaps that explains some of the weirdness seen in The Witcher 3, which bucks the trend of reasonable results seen under the DX11 API. Tahiti suffers badly here - in fact, the results are so poor at ultra settings that it doesn't matter if you're running at 1080p or 1440p as the result is essentially identical. At the higher resolution, the results seem to make more sense in comparison to later architectures.

We can't rule out driver issues - and one of the benefits of keeping the Novigrad test in our benchmark suite for so long is that we can remember test results in prior reviews where AMD hardware seemed to exhibit various issues, presumably sorted in later driver updates. However, with that said, ultra settings really pushes geometry, which is an area where GCN 1.0 was considerably off-pace. Could it really explain results this bad though? By the way, even under ultra, VRAM utilisation at both resolutions barely touches 2GB, so we're definitely not limited by memory here.

When The Witcher 3 launched, DF staffer Tom Morgan put together a list of PC settings that aimed to replicate the console experience - the idea being that in optimising the game for PC users, we could mirror the 'best bang for the buck' presets that CD Projekt RED chose for the game. The background character setting drops to low, water and texture quality adjust downwards to high while shadow quality, terrain quality and grass density are matched at medium. So how does this impact the performance differential? Well, 1080p performance is considerably improved on GCN 1.0 and the gap closes with the higher-end cards, producing a more plausible result, but there's still the sense that Tahiti should be performing better than it is here - and we can't rule out driver issues.

Additional tests on this page also throw up further curiosities. Take Forza Horizon 4, for example. This game was chosen for testing because it's a game where AMD graphics hardware has massively outperformed Nvidia equivalents. There's a decent likelihood that console optimisations for AMD graphics hardware may port over very nicely to modern Radeon GPUs via DX12, which would help to explain why it runs so well. However, this driver-level support doesn't seem to extend to GCN 1.0, where Tahiti delivers its most abysmal results yet. Again, 1080p performance in particular is way off pace. Bearing in mind how well Forza Horizon 4 runs on consoles using very similar Radeon technology, it's another test that fails to produce any convincing results.

Our final problem child is Battlefield 1 - another game that historically runs brilliantly on AMD hardware. Unlike Forza Horizon 4, Frostbite has support for both DX11 and DX12, allowing us to run the same content through both graphics APIs. When we run BF1 under DX11, there's a 17 per cent improvement over DX12 at 1080p, rising to 21 per cent at 1440p. It's interesting to note that when these new results are stacked up against Polaris and Navi, the performance increase from the newer architectures falls more into line with the results seen on our earlier DX11 testing. Cumulative percentage gains vs the DX11 result see a 55 per cent boost at 1080p, increasing to 58 per cent at 1440p. Once again, it's a result that suggests that AMD's focus on DX12 is mostly on products in its current line-up.

AMD RDNA vs GCN Analysis:

• Introduction, video analysis, synthetic benchmarks [This Page]
• Gaming benchmarks DX11: AC Unity, Crysis 3, Ghost Recon Wildlands, Far Cry 5
• Gaming benchmarks DX12: Rise/Shadow of the Tomb Raider, Strange Brigade, Wolfenstein 2
• Gaming benchmark problem children: Battlefield 1, Forza Horizon 4, The Witcher 3 [This Page]
• AMD architecture: Navi memory bandwidth scaling and conclusion