POWER9 Benchmarks vs. Intel Xeon vs. AMD EPYC Performance On Debian Linux

[xpue]

Power machine has 2x less cores

And still horribly loses in single-core benchmarks. Absolutely pathetic.

[George99]

And still horribly loses in single-core benchmarks. Absolutely pathetic.

I did a comparison with my ten year old AMD Athlon 7750 dual core processor. Surprisingly it has beaten the POWER9 in some benchmarks:

https://openbenchmarking.org/result/1805207-FO-1804049AR63

Michael I think the Parboil v2.5 Test: OpenMP MRI Gridding benchmark is broken. This result for my ancient artifact just can't be true!

[mongo]

It would be nice if there was a test of the quad precision floats as this is a first in hardware in this class of machine. As this is something that hasn't been delivered in x86 yet was similar to the capabilities of System/370-85 and some VAX machines it could be a huge win for some needs. Especially seeing as the SMID instructions don't even offer native support for the x87 80bit versions.

I doubt it will be important in the popular topics like ML today, as google's TPU actually converted 32 bit floats to 8 bits for improved performance.

But for some use cases it should be 100s or 1000s times faster than multiple-precision arithmetic methods used other places.

[uid_0]

These don't look like they are 8-core Power9 processors, they look like they're 4-core's. power consumption is too low. Is there a screenshot of the system specs?

[uid_0]

these don't look like they are 8-core Power9's, the power consumption is too low. Are they actually the 4-core Power9's? Can you post a screenshot of the system specs.

[azdaha]

Here are some interesting impressions of this great machine:

http://tenfourfox.blogspot.com/2018/04/unboxing-talos-ii-its-here.html

http://tenfourfox.blogspot.com/2018/...-talos-ii.html

Thank you for this

[Spooktra]

To expand on some of the other comments made in relation to the test results, all the apps tested were coded from the ground up with the x86 ISA in mind, on top of that some of them use hand crafted assembler and SIMD optimizations, that of course result in the Power9 systems getting blown away.

If you ever get access to such a system again I suggest you build your test suite on the Power9 system using whatever the most aggressive optimization settings are available for that compiler and then run the test suite.

I still think the x86 systems will beat it for the reasons outlined above but I will say this, both Intel and AMD are very lucky that IBM has no interest in the desktop market or that Apple doesn't wish to ink a deal with IBM and use their processors Apple's computers because if Apple released a Power Mac with a Power9 processor and then used it's money and resources to get optimized software for it;s new platform they would bury both Intel and AMD.

Power9's come in either SMT4 or SMT8 configurations, imagine a 12C96T based system with a BSD based optimized OS and optimized video/audio/photo editing apps and games.

[MarkJ]

IBM have an AC922 server based on the Power9. It might be a good comparison of specs/pricing. It would seem the AC922 is $5100(USD) and then there is the CPU at $2999 or $3999 (USD) depending on how many cores one wants. I presume thats each. Details taken from the IBM site tonight.

[Kango_V]

And then there is POWER10. 4 CPUs with 48 cores each and 8 threads per core. That's 1,536 threads. I know it would be expensive, power hungry and be in a huuuge box. But I'd like to see what that could do. Interestingly, POWER10 will include NVLink from nVidia. Hmmm.

[WorBlux]

And still horribly loses in single-core benchmarks. Absolutely pathetic.

That was never the point of these chips. Look at where they win, OS primatives, filesystem, and memory. It's a bandwidth optimized platform with a lot of throughput in a single address space. It's a more open platform and well situated to integrate with bespoke accelerator co-processors.

Though I suspect optimized scientific code could run fairly well. The cores are significantly wider than Skylake/Zen can have more store/loads in flight and a much cleaner concurrency model (not hard to do), but the OoO logic is much less deep, meaning performance is more sensitive to the compiler choices, cache misses, and memory layout.

And single thread is still useful, say compared to the newer SPAC layouts which have gone over to full database layout.