AMD Kaveri OpenCL Compared To Radeon & GeForce GPUs On Linux
Phoronix: AMD Kaveri OpenCL Compared To Radeon & GeForce GPUs On Linux
Earlier this week I delivered a wide range of NVIDIA GeForce and AMD Radeon graphics card tests from Ubuntu Linux and the focal point was the tests being done from the new AMD A10-7850K "Kaveri" APU. That testing found NVIDIA is leading over AMD with their binary graphics driver (of course, the same can't be said with AMD's superior open-source driver as yesterday's data showed), but how is the Linux OpenCL performance comparing between drivers and hardware? Here's the same set of NVIDIA and AMD graphics cards being benchmarked under Ubuntu with now looking at the OpenCL performance.
Since the APU Kaveri only crossfires with the R7 series, how about you test that against the AMD FX-8350 joined with the Radeon R-290 series and test the collaborative benefits of pre-HSA FX/R9 and the benefits of new HSA on an APU, while making note that until Excavator, you won't be able to leverage GCN 2.x GPGPUs with an APU?
bit/lite coin mining speed as OpenCL benchmark
There are a number of OpenCL-based bitcoin and litecoin mining programs, and typically, these programs are report the number of hashes/second, so it should not be too difficult to transform them into benchmarks. These might be nice additions, I guess there would be interest in mining in the community. (and maybe all this benchmarking might make you a negligible amount of money!)
JFYI, amazing Luxmark result is not related to actual hardware/driver/compiler state, but opposite. It is in fact result of hard work of Dade, Luxrender developer, who believe in AMD and for many years tune every bit of code to workaround all that AMD shitty compiler pitfalls. Imagine what he would have managed to do if AMD OpenCL compiler were open and enhanced same time, as GCC or even overrated APPLE'S RPOPRIETERY LAPDOG llvm that judging last new paid headlines already can compile before you even start make and turn your poor i5 into 16 core xeon cause it "better then GCC, believe me".
LLVM, is far beyond apple now. And let me guess, you have never used it!
Originally Posted by storm_st
Great tests but caution on intepreting the results.
Kudos for covering so much on Kaveri so soon especially on Linux where I think it matters much much more than Windows going forward. With Value and their SteamOS going full blast on Linux, we will see great developments ahead for the platform and specifically for Kaveri in particular mainly becuase of HSA. I have waited 25 years for HSA and would keep a close watch in the next few months on software and OS patches that will bring promise to heterogeneous computing for the home.
These tests based on current OpenCL 1.x is just indicative of the possibility of HSA and should not be taken as any sort of baseline in performance capability of HSA systems. The real test would be OS patches to the kernel for HSA and the corresponding drivers for the gpus and compute cores, then Open CL 2.0 layered over it. That is not enough as the apps needs to be OpenCL 2.0 complaint to allow the complete process to happen. So both those steps are not really here and hopefully coming in the next 2-3 quarters. The main OS players are not likely to do major changes to the kernel except for Linux. This is where I hold great hope for SteamOS in the gaming space and Ubuntu in the general use space.
For other people to talk about HSA without mentioning OpenCL 2.0 is just missing the software component. ie HSA means nothing if your apps cannot use it. SO as software develops for Kaveri, it would be great to see benchmark comparisons between HSA and non-HSA systems of comparable compute/gpu aggregate power. It should give people an indication of where they wanted their future build to focus on. It is also great to see AMD stepping up to improving the quality and speed of their drivers for Linux. It has been a long time. Rdgs.