AMD Kaveri APU performance

[juanrga]

I am preparing an article that estimates the performance of Kaveri next APU.

According to different sources, AMD official claims, rumours... I finally believe that the CPU on the A10 desktop Kaveri APU will be about a 25% faster than trinity A10-5800k. I know that this is approximate and that some benchmarks will show more performance whereas others will show much less, but provides an idea of performance.

I am using linux benchmarks from the review of the A10-5800k and adding a 25% performance. For instance, in the x264 kaveri would be on pair with the i5-2500k whereas that in John the Ripper would be faster than the i5-3470, but slower than both i5 in the linux kernel compilation benchmark.

I have several doubts, among them:

i)
Those tests use GCC 4.7. I know that 4.8.1 provides some performance improvements for both Intel and AMD. Future version of the compilers would offer additional performance gains (and maybe some regression). However, would we wait additional performance gains from the new "bdver3" flag? Or would be safe to consider that AMD's claims about performance already include the effects from new compiler? I believe that AMD claims are focused to Windows systems, where software will be not recompiled and that we will see extra performance gain under linux with recompilation, but I don't know if this has been true in the past.

ii)
Kaveri will be the first fully enabled HSA APU. AMD claims that developers are finding 5x improvement in the performance with HSA enabled software. Will be HSA enabled in linux with recompilation using the "bdver3" flag or HSA has to be enabled in the source code (rewrite the application)?

iii)
Will be the Phoronix tests suite prepared for HSA?

[ciplogic]

I am preparing an article that estimates the performance of Kaveri next APU.

According to different sources, AMD official claims, rumours... I finally believe that the CPU on the A10 desktop Kaveri APU will be about a 25% faster than trinity A10-5800k. I know that this is approximate and that some benchmarks will show more performance whereas others will show much less, but provides an idea of performance.

I am using linux benchmarks from the review of the A10-5800k and adding a 25% performance. For instance, in the x264 kaveri would be on pair with the i5-2500k whereas that in John the Ripper would be faster than the i5-3470, but slower than both i5 in the linux kernel compilation benchmark.

I have several doubts, among them:

i)
Those tests use GCC 4.7. I know that 4.8.1 provides some performance improvements for both Intel and AMD. Future version of the compilers would offer additional performance gains (and maybe some regression). However, would we wait additional performance gains from the new "bdver3" flag? Or would be safe to consider that AMD's claims about performance already include the effects from new compiler? I believe that AMD claims are focused to Windows systems, where software will be not recompiled and that we will see extra performance gain under linux with recompilation, but I don't know if this has been true in the past.

ii)
Kaveri will be the first fully enabled HSA APU. AMD claims that developers are finding 5x improvement in the performance with HSA enabled software. Will be HSA enabled in linux with recompilation using the "bdver3" flag or HSA has to be enabled in the source code (rewrite the application)?

iii)
Will be the Phoronix tests suite prepared for HSA?

In most cases the claims of 20+% claims should be taken with a grain of salt. The stated percent should be without any recompilation. This is theoretically possible by tuning the cache systems and branch predictors of the CPU.

So for your points GCC can improve the performance for both current CPUs and the future ones. Being said this, most GCC like optimizations that you get in GCC series 4.8 are mostly regarding LTO and both Linux Kernel and your typical programs are not compiled with it, so frankly you will see little improvement.

About HSA matters just in OpenCL or similar DirectX API, and it matters because your OpenCL code will not do memory copying from video card assigned memory to CPU and back. So you may see speed ups around image processing in some OpenCL filters in Gimp.

As far as I know HSA and OpenCL are not a part of Phoronox testing for many reasons, but at last the improvements may be noticed in in some years in OSS world, when OpenCL is easy to program, is not buggy on Amd hardware, etc.

At last Kaveri may come with higher efficiency but with lower frequencies, and a bit higher priced, so may nullify all advantages around performance, but we will see.

[juanrga]

In most cases the claims of 20+% claims should be taken with a grain of salt. The stated percent should be without any recompilation. This is theoretically possible by tuning the cache systems and branch predictors of the CPU.

So for your points GCC can improve the performance for both current CPUs and the future ones. Being said this, most GCC like optimizations that you get in GCC series 4.8 are mostly regarding LTO and both Linux Kernel and your typical programs are not compiled with it, so frankly you will see little improvement.

About HSA matters just in OpenCL or similar DirectX API, and it matters because your OpenCL code will not do memory copying from video card assigned memory to CPU and back. So you may see speed ups around image processing in some OpenCL filters in Gimp.

As far as I know HSA and OpenCL are not a part of Phoronox testing for many reasons, but at last the improvements may be noticed in in some years in OSS world, when OpenCL is easy to program, is not buggy on Amd hardware, etc.

At last Kaveri may come with higher efficiency but with lower frequencies, and a bit higher priced, so may nullify all advantages around performance, but we will see.

Thanks by useful replies. AMD is collaborating with LibreOffice to enable HSA. I expect HSA acceleration in spreadsheets as well.

[juanrga]

I have just uploaded a draft about what to expect from the next Kaveri APU. In general it is competing with Intel i5s. Much faster than IB i5 under GraphicsMagic, but much slower than a SB i5 under Himeno Benchmark

http://juanrga.com/en/AMD-kaveri-benchmark.html

I will update/correct when new info was available.