16-Way OpenCL Compute Comparison Of The Latest Polaris & Pascal GPUs

[phoronix]

Phoronix: 16-Way OpenCL Compute Comparison Of The Latest Polaris & Pascal GPUs

After running many OpenGL and Vulkan NVIDIA vs. AMD Linux benchmarks earlier this week, here is a 16-way graphics card comparison when testing the AMD Radeon "Polaris" and NVIDIA GeForce "Pascal" GPUs, among others, on Ubuntu 16.04 LTS and looking squarely at the OpenCL compute performance. Many OpenCL tests plus performance-per-Watt metrics too when using the latest NVIDIA proprietary Linux driver and AMDGPU-PRO.

http://www.phoronix.com/vr.php?view=23447

[Dedale]

Extremely instructive and handy. Thank you.

[chuckula]

Good testing Mike.

Quick question: Aside from SHOC, are these benchmarks representative of real-world OpenCL programs that people actually use or are they more just synthetic benchmarks?

[Mangix]

I suggest using hashcat for the OpenCL comparison as it's something people actually use. The -b parameter is helpful.

[LinuxID10T]

Good testing Mike.

Quick question: Aside from SHOC, are these benchmarks representative of real-world OpenCL programs that people actually use or are they more just synthetic benchmarks?

LuxMark is a render benchmark based off a real world application LuxRender. I'd say that is pretty real world.

[wizard69]

Good testing Mike.

Quick question: Aside from SHOC, are these benchmarks representative of real-world OpenCL programs that people actually use or are they more just synthetic benchmarks?

What difference does it make? Generally a specific programmers skills with respect to parallel programming makes a bigger difference. Synthetic benchmarks are actually very useful in their own right and like all benchmarks are only useful if the person using them is fully informed.

[wizard69]

Micheal;

very interesting and and a big thank you. A follow up hopefully uncovering the DP issues AMD is having would be welcomed. I find the power usage profile of AMD's new cards to be especially interesting. It will be great to see how NVidia fares with its new generation of cards. Also open driver results when they are capable would be most welcome also.

[Veerappan]

What difference does it make? Generally a specific programmers skills with respect to parallel programming makes a bigger difference. Synthetic benchmarks are actually very useful in their own right and like all benchmarks are only useful if the person using them is fully informed.

Agreed. And I'll add that performance tuning CL code for various architectures is pretty involved stuff. What benefits a given architecture will very often tank performance on another architecture.

[edwaleni]

Thanks for this OpenCL comparo, very helpful. I checked into a vendor that relies on GPU work and they said the Pascal's are not getting the results they expected. They won't put em in until they can find out where their app is not working well. Until then they are sticking with Maxwell's.

[TeoLinuX]

Am I losing something or AMDs get slaughtered here?
Are those benchmarks based on something created by/for nVidia, drivers are still very immature on AMD side?
No trolling: I'm no nVidia fan , I like OSS drivers (and I just have an intel with HD4000)