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

[torsionbar28]

Maybe on some top of the top, but when you look at Top 500 situation does not looks like that

I was talking about the top 10. Yes, the "top of the top". I.e. the newest and largest systems, and therefore the most relevant.

Some clunky old x86 supercomputer installed at the University of Tajikistan 12 years ago is of no importance here.

[ThinkOpenly]

Is there a reason why Rodinia is only '-O2' (not '-O3' like everything else), and for 7Zip, it seems no compile optimization at all? (Also, to make best use of the POWER9 processor, use '-mcpu=power9').

Disclaimer: I work for IBM, but do not speak for them.

[Thovthe]

I would be very interested in seeing these in future POWER articles:[LIST][*]Comparisons where as much as possible is kept the same between systems.[*]Comparisons with similar core counts.[*]A discussion of what the experience is like with day to day desktop use on POWER9. Maybe debian? They seem to have the best support for PPC64.[*]Maybe a look at what one would need to be prepared to do if one wanted to move to POWER for day to day because they were interested in computing freedom.[*]Big ask: Reducing the SMT to SMT2 on the Talos and overclocking to see what that does to single threaded workloads.

A couple more interesting configurations occur to me:

• Turning off cores on the Talos so that one can make a more direct comparison between these platforms.
• Disabling/removing one CPU and the associated memory on the Talos II. Comparing this single socket system to an equal core count and equal total memory configured Talos II to illustrate differences between on die core number to performance scaling and dual socket core number to performance scaling.

[Thovthe]

This comparison (also the first one with the HDD) seems to be needlessly tipped by differences between hardware in the test systems. In the first set you compared x86 systems with SSDs to a Talos II with an HDD; I know the Talos had to have an HDD but surely you could have stuck an HDD in the compared systems. Now you have different storage drives between the x86 and POWER as well as ~half as much ram on the x86 systems compared to the Talos. Seems like there's little that can be said concretely with this being the case. Is there some reason for these differences that I am not understanding?

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

I would be very interested in seeing these in future POWER articles:

• Comparisons where as much as possible is kept the same between systems.
• Comparisons with similar core counts.
• A discussion of what the experience is like with day to day desktop use on POWER9. Maybe debian? They seem to have the best support for PPC64.
• Maybe a look at what one would need to be prepared to do if one wanted to move to POWER for day to day because they were interested in computing freedom.
• Big ask: Reducing the SMT to SMT2 on the Talos and overclocking to see what that does to single threaded workloads.

I know that this post is quite critical but I appreciate what you do here phoronix Michael

[classichasclass]

My recommendation is HQEMU -- it has POWER support, and is fast compared to QEMU TCG. Even though system mode is a little buggy, it worked enough to boot an XP disk and the difference was stark -- "unusably slow" (normal QEMU TCG) vs. "this PC just feels a little underpowered for XP" (HQEMU). User mode is supposed to work significantly better, but I haven't had a chance to try it out yet.

http://itanium.iis.sinica.edu.tw/hqemu/

Tim, is this new? I remember looking at the last version and saying "would be nice if there were Power support" but it looks like this was recently added.

Cameron Kaiser

[dungeon]

You can't measure just performance share, because that is relative as there usually newest are better Of these on Top 10 eldest is just 6 years old

That is like saying, only Top 10 games on Steam, Playstation, etc... are worthwhile and everything else is crap

Or like to say - only these currently play tennis and everything else is crap Of course there is some truth into this, represent just a moment in time, so relative.

[frief]

Hi dungeon,

the above diagram shows ".. Processor Family" by "Number of Systems".
This means that the slowest system on the list with its 838.900.000.000.000 floating point operations per second is shown with the same weight as the fastest system.

In the context here probably "Processor Family" by "Performance Share" would be more adequate.
But even that would likely not be completely fair because GPUs/accelerators would not have been taken into account.

[dungeon]

Lol, what x86 dominance in super computers? Only 4 of the top 10 largest supercomputers in the world are x86. The x86 supercomputer fad died in the last decade.

The top supercomputers nowadays are all based on POWER, weird Chinese chips, or Xeon Phi cards.

Maybe on some top of the top, but when you look at Top 500 situation does not looks like that



These Top 500 are spreaded across 25 countries:

https://en.wikipedia.org/wiki/TOP500#Top_countries

or maybe we could simplify this... so to say:

1. CHN 206
2. RoW (Rest of the World) 170
3. USA 124

[rabcor]

Man that thing is a monster, it's doing more than just keeping up... Only downside seemingly is the disproportionately increased power use, 100w... but it's not that big of a deal unless you're running an entire server farm of the stuff.

I kinda wish they'd make consumer class CPUs to compete with ryzen and the core-i series though. I'll take higher power use if it means an open source CPU that can compete this narrowly with AMD and Intel.

[WolfpackN64]

My recommendation is HQEMU -- it has POWER support, and is fast compared to QEMU TCG. Even though system mode is a little buggy, it worked enough to boot an XP disk and the difference was stark -- "unusably slow" (normal QEMU TCG) vs. "this PC just feels a little underpowered for XP" (HQEMU). User mode is supposed to work significantly better, but I haven't had a chance to try it out yet.

http://itanium.iis.sinica.edu.tw/hqemu/

Well, until now I didn't know there were other options still available. Might be interesting to see which one can achieve the best performance in binary translation.