An Initial Look At The IBM POWER9 4-Core / 16-Thread CPU Performance On The Blackbird

[jacob]

IDK if it's assembler optimizations or just not using the appropriate operation or at the appropriate sizes

It could be simply that these graphics/audio apps use SSE or AVX but fall back on default scalar code on the POWER9. If that's the case then adding Altivec support would make a huge difference (and it would be a nice little project too).

[edwaleni]

Get VMX/VSX/VMX128 support into the applications and I would imagine the results will change dramatically. Especially for Dav1D, Zip and other tests that can take advantage of the SIMD and vector capabilities.

[wizard69]

Very interesting results. I’m wondering though if some of these numbers may indicate less than optimal results from running 4 threads on one core. We have seen instance where turning off threading per core can increase performance on X86 so might we also see a bigger hit from the odd bit of code on a 4 threads per core processor. That is some applications trashing threading on these cores?

[andreano]

I also liked the commentary and summary. Michael is often pretty short on words in benchmarks, but I do appreciate his explanatory notes.

[Guest]

This looks pretty interesting, and it shows the potential of POWER workstations. While software support seems to be a bit rough around the edges, it could be a lot worse. Considering how much freedom one is getting with these chips, I think they're worth considering, especially considering that buying them supports efforts of making POWER viable for this sort of use.

[skeevy420]

Very interesting results. I’m wondering though if some of these numbers may indicate less than optimal results from running 4 threads on one core. We have seen instance where turning off threading per core can increase performance on X86 so might we also see a bigger hit from the odd bit of code on a 4 threads per core processor. That is some applications trashing threading on these cores?

I was wondering the same thing and if it could also be cpu schedulers, etc not being as optimized for the 4:1 thread to core ratio used here over the more common 2:1 ratio that the average x86 desktop and server use.

[skeevy420]

This looks pretty interesting, and it shows the potential of POWER workstations. While software support seems to be a bit rough around the edges, it could be a lot worse. Considering how much freedom one is getting with these chips, I think they're worth considering, especially considering that buying them supports efforts of making POWER viable for this sort of use.

Especially with these lower priced parts to get more enthusiasts on board.

It'll be really interesting to see how an RX 570 or better AMD card will be able to run Wine games in another year or two on one of these.

EDIT: Or one of those Intel GPUs since they're likely going to be open source. I think it would be neat & very ironic to play x86 Windows games on a POWER9 CPU with an Intel GPU running some version of Linux. That would have been a very crackheaded idea just a couple of years ago.

[RavFX]

Especially with these lower priced parts to get more enthusiasts on board.

It'll be really interesting to see how an RX 570 or better AMD card will be able to run Wine games in another year or two on one of these.

EDIT: Or one of those Intel GPUs since they're likely going to be open source. I think it would be neat & very ironic to play x86 Windows games on a POWER9 CPU with an Intel GPU running some version of Linux. That would have been a very crackheaded idea just a couple of years ago.

Forget about Wine game. Wine is not an emulator.Code is passed straight to the CPU except Windows ABI call ar converted to equivalent Linux one. Power9 is not going to execute X86/AMD64 code.

Still, I would be buying one if it was not like 3KUS to get a ATX-equivalent board for the number of usable PCI-e slot.

[sireangelus]

i think that the price is way too high for the performance given

[Snaipersky]

Forget about Wine game. Wine is not an emulator.Code is passed straight to the CPU except Windows ABI call ar converted to equivalent Linux one. Power9 is not going to execute X86/AMD64 code.

Still, I would be buying one if it was not like 3KUS to get a ATX-equivalent board for the number of usable PCI-e slot.

You forget Hangover. There's still work to be done with regard to thunking I believe, but as long as the processor is run in LEM, then it should work to some degree.