Hands On With The Most Open-Source, High-Performance System For 2018

[madscientist159]

Yep will be running POWER tests at various Spectre levles.

And yes the defconfig does differ for Linux. As stated, this was basically sanity check after I got Ubuntu freshly installed and just wanted to fire off something to verify no immediate stability or thermal problems.

Sure, sounds great! Just wanted to mention that in case anyone was tempted to directly compare the raw kernel compile time number against some other box.

[edwaleni]

Following the Chromium for POWER thread, the biggest issue for any POWER based browser is going to be media (codec) handling. Not all of the compiles are aware of SSE/VMX/VSX/VMX128 for POWER and this causes the media to either not play or play with choppiness. I know Talos is working with IBM on improving this. Hopefully a solution has been found as it seems many want to use a DE with this.

[SofS]

Echoing some requests here, it would be good to see how different compilers handle native optimizations. At least gcc and clang with and without the native target. Other languages besides C like C++, Haskell, Fortran and Ada would be good additions. If possible also how well the increase in build parallelism scales (make -j#).

[dimtass]

I would be interested to know in which kind of industry a machine like that is used and a real case scenario.

[kgardas]

Here's a few I ran just as some sanity checks this morning - https://openbenchmarking.org/result/...SK-POWER9COM35

On LLVM you are touching speed of 2x Xeon Gold 6138 which is great. For the less price I guess same perf + completely open system! Not sure if RAM/storage was comparable though. BTW, with all those thread on POWER9 I'm afraid LLVM compile benchmark does not scale wall due to linking step.

Thanks a lot for this first teaser, looking forward to see more test of this wonderful machine!

[kgardas]

Michael, btw, in the article you mention single socket power board for $2k, but this is not true, for this price you can have bundle of cpu/heatsing and talos lite, so please fix this wording a bit not to confuse interested readers. Lite (board only) is still for just $1.1k Thanks!

[pegasus]

From the photos it seems that only half of the memory channels are populated? IIRC P9 also has 8 ddr channels?

So before you do any serious benchmarking, get the memory sorted out first.

[Adarion]

Would like to know about compilation performance and idle power consumption.

What he said plus some comparison to recent x86/amd64 systems of this class.
I'm getting slightly envious when I see that big fat box of libre computing power.

[Dawn]

From the photos it seems that only half of the memory channels are populated? IIRC P9 also has 8 ddr channels?

So before you do any serious benchmarking, get the memory sorted out first.

Sforza, the SCM used by these machines, has four channels. Monza and LaGrange, the other OpenPower SCM options, have eight (but a bigger socket.)

[pegasus]

So direct comparison would be threadripper, not epyc. Ok, noted, thanks.