Raptor Talos II POWER9 Benchmarks Against AMD Threadripper & Intel Core i9

[Maddo]

Ehm, unless I missed something, aren't these result kinda bogus since we're comparing a dual socket setup against a single socket one?

[madscientist159]

Does Talos II support off the shelf AMD/Nvidia graphics cards or the card needs special BIOS/firmware that contains POWER compatbile binary? My SPARC64 VII for instance needs special Radeon card that has BIOS compiled for SPARC arch.

If there are open drivers for the card, it just works for the most part. For using the GPU in petitboot (before the target operating system has loaded) the only obstacle is proprietary GPU firmware, which we do not ship as part of the base Talos II firmware image (unless a specific GPU was ordered as part of a built system).

We've been working on ways to simplify providing the proprietary GPU firmware to the bootloader, see for instance https://wiki.raptorcs.com/wiki/Talos..._To_BOOTKERNFW, but there's still some work to go (including making this as simple as point + click). Long term it would be great if AMD would provide open firmware for their GPUs and compute cards, so that either it could be flashed to the GPU itself or shipped as part of the base firmware image, but AMD has publicly stated they cannot do so at this time.

If you don't want to mess around with GPU firmware and the bootloader, you don't need to -- just use the onboard VGA or serial to interact with the bootloader and allow the OS to load the proprietary firmware for your GPU.

[hax0r]

Does Talos II support off the shelf AMD/Nvidia graphics cards or the card needs special BIOS/firmware that contains POWER compatbile binary? My SPARC64 VII for instance needs special Radeon card that has BIOS compiled for SPARC arch.

[Michael]

How long will you keep the system? Would love to see more and more tests.

Presumably for as long as the hardware remains relevant and still being used for benchmarking, there isn't any set date.

[Yoshi]

Really impressive results. Hopefully RaptorCS continues to be successful.

[darkbasic]

I will have more tests like those (SMT on/off, Spectre tests, etc) coming soon. Have had the system now less than a week

How long will you keep the system? Would love to see more and more tests.

[carewolf]

Having such a strong completely open system is pretty attractive. Thanks for the article.

However, since Power 9 systems offer up to 8 threads per core (seemingly 4 in this case) and the Portsmash security vulnerability seems to affect all SMT CPUs I would have been interested how the system performs without SMT.
I would, due to the new cognitions, be interested in comparisons between enabled and disabled SMT for CPUs in general when possible.
I read about several statements of people working closer with the hardware that SMT itself might perhaps never be a safe technique.

Powersmash was a timing bug in OpenSSL it has already been fixed. Stop freaking out about it. Any program that is vulnerable to portsmash like attacks is also vulnerable to timing attacks, so it programs just have to be written with out branching code paths based on secret data.

[oooverclocker]

I will have more tests like those (SMT on/off, Spectre tests, etc) coming soon. Have had the system now less than a week

Cool, thank you for the quick reply!

[Michael]

Having such a strong completely open system is pretty attractive. Thanks for the article.

However, since Power 9 systems offer up to 8 threads per core (4 in this case) and the Portsmash security vulnerability seems to affect all SMT CPUs I would have been interested how the system performs without SMT.
I would, due to the new cognitions, be interested in comparisons between enabled and disabled SMT for CPUs in general when possible.
I read about several statements of people working closer with the hardware that SMT itself might perhaps never be a safe technique.

I will have more tests like those (SMT on/off, Spectre tests, etc) coming soon. Have had the system now less than a week

[oooverclocker]

Having such a strong completely open system is pretty attractive. Thanks for the article.

However, since Power 9 systems offer up to 8 threads per core (seemingly 4 in this case) and the Portsmash security vulnerability seems to affect all SMT CPUs I would have been interested how the system performs without SMT.
I would, due to the new cognitions, be interested in comparisons between enabled and disabled SMT for CPUs in general when possible.
I read about several statements of people working closer with the hardware that SMT itself might perhaps never be a safe technique.