Compute

Shifting gears, we have our look at compute performance. As an FP64 card, the R9 Fury X only offers the bare minimum FP64 performance for a GCN product, so we won’t see anything great here. On the other hand with a theoretical FP32 performance of 8.6 TFLOPs, AMD could really clean house on our more regular FP32 workloads.

Starting us off for our look at compute is LuxMark3.0, the latest version of the official benchmark of LuxRender 2.0. LuxRender’s GPU-accelerated rendering mode is an OpenCL based ray tracer that forms a part of the larger LuxRender suite. Ray tracing has become a stronghold for GPUs in recent years as ray tracing maps well to GPU pipelines, allowing artists to render scenes much more quickly than with CPUs alone.

Compute: LuxMark 3.0 - Hotel

The results with LuxMark ended up being quite a bit of a surprise, and not for a good reason. Compute workloads are shader workloads, and these are workloads that should best illustrate the performance improvements of R9 Fury X over R9 290X. And yet while the R9 Fury X is the fastest single GPU AMD card, it’s only some 16% faster, a far cry from the 50%+ that it should be able to attain.

Right now I have no reason to doubt that the R9 Fury X is capable of utilizing all of its shaders. It just can’t do so very well with LuxMark. Given the fact that the R9 Fury X is first and foremost a gaming card, and OpenCL 1.x traction continues to be low, I am wondering whether we’re seeing a lack of OpenCL driver optimizations for Fiji.

For our second set of compute benchmarks we have CompuBench 1.5, the successor to CLBenchmark. CompuBench offers a wide array of different practical compute workloads, and we’ve decided to focus on face detection, optical flow modeling, and particle simulations.

Compute: CompuBench 1.5 - Face Detection

Compute: CompuBench 1.5 - Optical Flow

Compute: CompuBench 1.5 - Particle Simulation 64K

Quickly taking some of the air out of our driver theory, the R9 Fury X’s performance on CompuBench is quite a bit better, and much closer to what we’d expect given the hardware of the R9 Fury X. The Fury X only wins overall at Optical Flow, a somewhat memory-bandwidth heavy test that to no surprise favors AMD’s HBM additions, but otherwise the performance gains across all of these tests are 40-50%. Overall then the outcome over who wins is heavily test dependent, though this is nothing new.

Our 3rd compute benchmark is Sony Vegas Pro 13, an OpenGL and OpenCL video editing and authoring package. Vegas can use GPUs in a few different ways, the primary uses being to accelerate the video effects and compositing process itself, and in the video encoding step. With video encoding being increasingly offloaded to dedicated DSPs these days we’re focusing on the editing and compositing process, rendering to a low CPU overhead format (XDCAM EX). This specific test comes from Sony, and measures how long it takes to render a video.

Compute: Sony Vegas Pro 13 Video Render

At this point Vegas is becoming increasingly CPU-bound and will be due for replacement. The Fury X none the less shaves off an additional second of rendering time, bringing it down to 21 seconds.

Moving on, our 4th compute benchmark is FAHBench, the official Folding @ Home benchmark. Folding @ Home is the popular Stanford-backed research and distributed computing initiative that has work distributed to millions of volunteer computers over the internet, each of which is responsible for a tiny slice of a protein folding simulation. FAHBench can test both single precision and double precision floating point performance, with single precision being the most useful metric for most consumer cards due to their low double precision performance. Each precision has two modes, explicit and implicit, the difference being whether water atoms are included in the simulation, which adds quite a bit of work and overhead. This is another OpenCL test, utilizing the OpenCL path for FAHCore 17.

Compute: Folding @ Home: Explicit, Single Precision

Compute: Folding @ Home: Implicit, Single Precision

Compute: Folding @ Home: Explicit, Double Precision

Both of the FP32 tests for FAHBench show smaller than expected performance gains given the fact that the R9 Fury X has such a significant increase in compute resources and memory bandwidth. 25% and 34% respectively are still decent gains, but they’re smaller gains than anything we saw on CompuBench. This does lend a bit more support to our theory about driver optimizations, though FAHBench has not always scaled well with compute resources to begin with.

Meanwhile FP64 performance dives as expected. With a 1/16 rate it’s not nearly as bad as the GTX 900 series, but even the Radeon HD 7970 is beating the R9 Fury X here.

Wrapping things up, our final compute benchmark is an in-house project developed by our very own Dr. Ian Cutress. SystemCompute is our first C++ AMP benchmark, utilizing Microsoft’s simple C++ extensions to allow the easy use of GPU computing in C++ programs. SystemCompute in turn is a collection of benchmarks for several different fundamental compute algorithms, with the final score represented in points. DirectCompute is the compute backend for C++ AMP on Windows, so this forms our other DirectCompute test.

Compute: SystemCompute v0.5.7.2 C++ AMP Benchmark

Our C++ AMP benchmark is another case of decent, though not amazing, GPU compute performance gains. The R9 Fury X picks up 35% over the R9 290X. And in fact this is enough to vault it over NVIDIA’s cards to retake the top spot here, though not by a great amount.

Synthetics Power, Temperature, & Noise
Comments Locked

458 Comments

View All Comments

  • Navvie - Thursday, July 2, 2015 - link

    "Which is not say I’m looking" (paragraph 5, first line).

    Missing a "to" I think.
  • watzupken - Thursday, July 2, 2015 - link

    Brilliant review. Well worth the wait. Thanks Ryan.
  • Taracta - Thursday, July 2, 2015 - link

    ROPs, ROPs, ROPs! Hows can they ~ double everything else and keep the same amount of ROPs and expect to win?
  • Thatguy97 - Thursday, July 2, 2015 - link

    maybe something to do with cost or yield
  • tipoo - Thursday, July 2, 2015 - link

    They literally hit the size limits interposers can scale up to with this chip - so they can't make it any bigger to pack more transistors for more ROPs, until a die shrink. So they decided on a tradeoff, favouring other things than ROPs.
  • Kevin G - Thursday, July 2, 2015 - link

    They had a monster shader count and likely would be fine if they went to 3840 max to make room for more ROPs. 96 or 128 ROPs would have been impressive and really made this chip push lots of pixels. With HBM and the new delta color compression algorithm, there should be enough bandwidth to support these additional ROPs without bottle necking them.

    AMD also scaled the number of TMUs with the shaders but it likely wouldn't have hurt to have increased them by 50% too. Alternatively AMD could have redesigned the TMUs to have better 16 bit per channel texture support. Either of these changes would have put the texel throughput well beyond the GM200's theoretical throughput. I have a feeling that this is one of the bottlenecks that helps the GM200 pull ahead of Fiji.
  • tipoo - Friday, July 3, 2015 - link

    Not saying it was the best tradeoff - just explaining. They quite literally could not go bigger in this case.
  • testbug00 - Sunday, July 5, 2015 - link

    the performances scaling as resolution increase is better than Nvidia, implying the ROPs aren't the bottleneck...
  • chizow - Sunday, July 5, 2015 - link

    No, that implies the shaders are the bottleneck at higher resolutions while ROP/fillrate/geometry remained constant. While Nvidia's bottleneck at lower resolutions isn't shader bound but their higher ROP/fillrate allows them to realize this benefit in actual FPS, AMD's ROPs are saturated and simply can't produce more frames.
  • Ryan Smith - Thursday, July 2, 2015 - link

    Right now there's not a lot of evidence for R9 Fury X being ROP limited. The performance we're seeing does not have any tell-tale signs of being ROP-bound, only hints here and there that may be the ROPs, or could just as well be the front-end.

    While Hawaii was due for the update, I'm not so sure we need to jump up in ROPs again so soon.

Log in

Don't have an account? Sign up now