AMD architecture performance analysis: DirectX12/Vulkan Focus

For those who've jumped straight to this page, a quick recap: game analysis was based on a PC running on a Core i7 8700K system running at stock frequencies, with games typically running all cores and threads at 4.3GHz. Accompanying that is system memory supplied by GSkill - 3200MHz CL14 FlareX modules - while our motherboard is an MSI Gaming X370 product. All games are run from SSD, with benchmarks calculated using FCAT data. Essentially, our chosen test titles are run with v-sync disabled at the settings described, and each new frame is marked up with a brightly coloured border. By tracking the border in our software, we can determine frame-rate and frame-time and overlay graphs with performance in context of the content. Multi-gigabyte video captures are condensed into frame data .txt files that are fed into the Eurogamer CMS to create the graphs below. Every single frame is therefore available, with frame-times viewable by pressing play on the appropriate YouTube insert.

For users running on PC or Mac, the graphs themselves are interactive. Mouse over a particular entry to get the appropriate reading, or better yet, click the graph for the arguably much more useful percentage differentials. In short, every frame we capture is represented on these pages and you have as much or as little performance data as you desire. At-a-glance or deep-dive, it's all here.

Our next collection of games is using the new wave of low-level graphics APIs, DirectX 12 and its open standards equivalent, Vulkan - an evolution of AMD's prior work on its own API, Mantle (remember that?). I was keen to try out a range of DX12 titles as AMD has been at the forefront of DX12 development and asynchronous compute in particular is a feature that can strongly benefit AMD hardware. What quickly becomes evident though is that GCN 1.0 isn't a particularly strong DX12 performer. Navi vs Tahiti comparisons here show stratospheric gains for the new architecture, but the inclusion of our Polaris vs Navi testing shows gains very much in like with our DX11 results. Interesting as they are, I don't think these results accurately chart the performance value of a GCN vs RDNA at the same core clocks, but let's dig in regardless.

Rise of the Tomb Raider

Our first look at the Foundation Engine from Crystal Dynamics highlights the problem. Incremental gains from Polaris to Navi sit nicely alongside the kind of results we gleaned from DX11 testing, but GCN 1.0 is way, way off pace here, regardless of the resolution you may choose. The notion of modern game engines struggling on older hardware is nothing new, of course, and I can confirm that VRAM usage at both resolutions is nowhere near Tahiti's 3GB limit. The lack of additional performance from Polaris's extra CUs in the Radeon RX 580 continues to disappoint, something that extends to most of the games on this page. Cumulative percentage gains across the architectures result in Navi delivering an 84 per cent increase over Tahiti at 1080p, dropping to 77.8 per cent at 1440p. It's a fairly implausible gap, and likely down to driver and API support.

Shadow of the Tomb Raider

The later iteration of the Foundation Engine shows Navi extending its lead over Polaris a touch, but gains over Tahiti are still in 75 to 88 per cent territory, depending on resolution. As we'll see on the next page, there are other limitations that may cause Tahiti to stall - such as geometry set-up, for example. However, here, we're running Shadow at medium settings, so this seems unlikely. More likely is that GCN 1.0's design predates DirectX 12 and its feature set is considerably sparser compared to Polaris and Navi.

Strange Brigade

Strange Brigade is one of AMD's favourites when it comes to benchmarking, with Rebellion's engine working extremely well with Radeon graphics hardware. The engine plays to the strengths of Team Red's GPUs and utilises asynchronous compute extensively - a godsend to AMD hardware which can be underutilised. Once again, GCN 1.0 via our Radeon R9 280X fails to perform adequately and gen-on-gen gains are likely exaggerated by API or driver limitations or perhaps even geometry processing bottlenecks in the older hardware. An interesting element of this title is that both Vulkan and DX12 are supported - but Vulkan is a touch slower on pretty much any GPU you care to test, and that includes the 280X.

Wolfenstein 2: The New Colossus

The lack of meaningful data from GCN 1.0 continues to be a source of frustration as id Tech 6 pushes the current generation of consoles hard - and the medium setting we chose here (to accommodate the 280X's VRAM limits) is essentially the equivalent of console visual quality. Also frustrating is that our graphing system fails to accurately render RX 570 results so we've had to omit them. Regardless, with the gap between GCN 1.0 and RDNA 1.0 proving so vast, we can only assume that Vulkan support or driver support for Tahiti is limited. Head-to-head gains for Navi vs Polaris look plausible - and in line with many of our DX11 results.

AMD RDNA vs GCN Analysis:

• Introduction, video analysis, synthetic benchmarks
• 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 [This Page]
• Gaming benchmark problem children: Battlefield 1, Forza Horizon 4, The Witcher 3
• AMD architecture: Navi memory bandwidth scaling and conclusion