13-Way Radeon AMDGPU-PRO 17.50 vs. NVIDIA Linux OpenCL Compute Comparison

[andrei_me]

Why is blender so good on the Vega cards, some optimization happened? bridgman

[bridgman]

I think it's mostly a function of the size of the work units that get pushed to the GPU. If they are fairly small you end up transfer-bound and maybe even CPU-bound... but if they are fairly large then the raw power of the GPU gets to show off.

Or why is RX560 faster than RX580 with blender?

I don't fully understand that yet. The easy answer is that the work units being pushed to GPU are "560-sized" and don't take full advantage of 580 so clock speed rules (smaller GPUs can often clock higher), but that doesn't explain what is happening with Vega. Vega has higher clocks but not high enough to account for the difference in results.

[nuetzel]

Why is blender so good on the Vega cards, some optimization happened? bridgman

Or why is RX560 faster than RX580 with blender?
BTW I've no luck running RHL packages on SLES/openSUSE Tumbleweed so far.
Hurry up!

[nuetzel]

I think it's mostly a function of the size of the work units that get pushed to the GPU. If they are fairly small you end up transfer-bound and maybe even CPU-bound... but if they are fairly large then the raw power of the GPU gets to show off.



I don't fully understand that yet. The easy answer is that the work units being pushed to GPU are "560-sized" and don't take full advantage of 580 so clock speed rules (smaller GPUs can often clock higher), but that doesn't explain what is happening with Vega. Vega has higher clocks but not high enough to account for the difference in results.

Mostly I agree, but by all means 'clock speed' 560 vs 580?

Chip: Polaris 21 XT (Baffin XT) "GCN Gen4" • Chiptakt: 1175MHz, Boost: 1176MHz • Speicher: 4GB GDDR5, 1750MHz, 128bit, 112GB/s • Shader-Einheiten/TMUs/ROPs: 1024/64/16 • Rechenleistung: 2408GFLOPS (Single), 150GFLOPS (Double) • Fertigung: 14nm • Leistungsaufnahme: 80W (TDP)

Chip: Polaris 20 XTX (Ellesmere XTX) "GCN Gen4" • Chiptakt: 1257MHz, Boost: 1411MHz • Speicher: 8GB GDDR5, 2000MHz, 256bit, 256GB/s • Shader-Einheiten/TMUs/ROPs: 2304/144/32 • Rechenleistung: 6502GFLOPS (Single), 406GFLOPS (Double) • Fertigung: 14nm • Leistungsaufnahme: 185W (TDP)

[hoohoo]

Good work, AMD!

[mantu]

Hmm, my 1080 ti renders the blender BMW27 scene in 120 seconds. On the first render it takes longer, because it AFAIK compiles the kernels. With CUDA rendering takes 50 seconds.

[MrCooper]

Then why was R9 Fury winning in 2 of the 3 tests?

Fury has slightly more memory bandwidth than Vega.

[thelongdivider]

Great benchmarking set!

Hmm, my 1080 ti renders the blender BMW27 scene in 120 seconds. On the first render it takes longer, because it AFAIK compiles the kernels. With CUDA rendering takes 50 seconds.

I think Nvidia cares very little about opencl. Cuda is the direction they want to go to keep their code private. It is very fast though for my purposes (on a 1070).

[mantu]

Great benchmarking set!



I think Nvidia cares very little about opencl. Cuda is the direction they want to go to keep their code private. It is very fast though for my purposes (on a 1070).

Yep, that's what I have heard too.

Btw, even a 970 renders the scene in 122 seconds with CUDA.