Path: ...!news.mixmin.net!eternal-september.org!feeder3.eternal-september.org!news.eternal-september.org!.POSTED!not-for-mail
From: Michael S <already5chosen@yahoo.com>
Newsgroups: comp.arch
Subject: Re: is Vax addressing sane today
Date: Sun, 6 Oct 2024 12:40:00 +0300
Organization: A noiseless patient Spider
Lines: 105
Message-ID: <20241006124000.00001325@yahoo.com>
References: <2024Oct3.085754@mips.complang.tuwien.ac.at>
	<memo.20241003234930.19028I@jgd.cix.co.uk>
	<2024Oct4.170717@mips.complang.tuwien.ac.at>
	<vdsmgb$ukl1$3@dont-email.me>
	<2024Oct6.091859@mips.complang.tuwien.ac.at>
MIME-Version: 1.0
Content-Type: text/plain; charset=US-ASCII
Content-Transfer-Encoding: 7bit
Injection-Date: Sun, 06 Oct 2024 11:39:29 +0200 (CEST)
Injection-Info: dont-email.me; posting-host="18ac7e79562038ddb85a7a321c10035e";
	logging-data="1251000"; mail-complaints-to="abuse@eternal-september.org";	posting-account="U2FsdGVkX19j7fVJ2SJ1zmmq4yTcoZeBWKxSrAETg5M="
Cancel-Lock: sha1:GkQfPFnNgrz5qQhZidhZ3RCY1hA=
X-Newsreader: Claws Mail 3.19.1 (GTK+ 2.24.33; x86_64-w64-mingw32)
Bytes: 5802

On Sun, 06 Oct 2024 07:18:59 GMT
anton@mips.complang.tuwien.ac.at (Anton Ertl) wrote:

> Lawrence D'Oliveiro <ldo@nz.invalid> writes:
> >Native POWER is, I think, called pSeries. It continues to sell in
> >its own right because it offers high performance--high enough to
> >earn a few ongoing spots near the top of the Top500 supercomputer
> >list.  
> 
> Looking at the June 2024 edition, I see Summit as the highest-ranked
> system with Power CPUs, and they are Power 9.  So if your claim was
> true that the Top500 supercomputer list reflects CPU performance,
> Power 9 would beat Power 10 in CPU performance, and EPYC, Xeon,
> Fujitsu A64FX and Nvidia Grace are more powerful CPUs.  However, in
> most supercomputers (including Summit) the GPGPUs provide the bulk of
> the FLOPS that are measured in the Top 500, so looking at the Top 500
> is misleading for determining CPU performance.
>

Yes, in almost all top entries in top500 the compute muscle is GPGPU,
with CPUs playing role of glorified Peripheral Processor of ancient
supercomputers. That applies both to POWER and to Xeons and to EPYC.

However there are two exceptions: Fugaku (#4, Fujitsu A64Fx) and Sunway
TaihuLight (#13, Sunway SW26010).

Majority of GPUs in the list are NVidia of various generations, but the
#1 (US DOE Frontier) uses AMD GPUs and #2 (US DOE Aurora) uses Intel
GPUs.

BTW, IBM Summit (NVidea GV100+IBM POWER9), despite still being pretty
high on the Top500 list, is going to be retired next month.
I wonder if Sunway TaihuLight is aging better.


> So let's look at SPEC CPU instead.  For CPU2017, I see only four
> entries from IBM, all for the Integer Rate metric, two with Power 9
> and two with Power 10 CPUs.  The highest of those results is:
> 
> base peak
> 1700 2170 IBM Power E1080
> 
> That's with 8 sockets, 120 cores, and 960 threads.  Looking at other
> 8-socket machines, I find
> 
> base peak
> 3820 3880 BullSequana SH80
> 
> That's with 8 sockets, 480 cores, and 960 threads (similar results
> from Fujitsu PRIMERGY RX8770 M7, HPE Compute Scale-up Server 3200,
> Inspur TS860G7 and Supermicro SuperServer SYS-681E-TR, all done with
> Xeon Platinum 8490H CPUs).  And if you go for maximum performance,
> there's a 16-socket Xeon machine from Bull with base=7400, peak=7450.
> 
> Alternatively, you can instead buy a 2-socket system with similar
> performance to the 8-socket IBM Power E1080:
> 
> base peak
> 1950 2140 ASUS RS720A-E12-RS12
> 
> and similar results from other systems with the EPYC 9754.
> 
> https://www.spec.org/cpu2017/results/res2021q3/cpu2017-20210814-28679.html
> https://www.spec.org/cpu2017/results/res2024q3/cpu2017-20240701-43944.html
> https://www.spec.org/cpu2017/results/res2023q2/cpu2017-20230522-36617.html
> 
> Admittedly, IBM extracts the most performance from each core, but with
> only 15 cores per CPU (where others have 128), that is no longer that
> impressive. 

"Core" in POWER9 is sort-of cheating. For nearly all practical purposes
what they call 'core' is a couple of cores with just a little bit of
resource sharing between halves when running in single-thread mode.
Just enough to have a judicial justification to being called 'core'.
I don't know if POWER10 is similar or different in that regard.

> Nevertheless, neither machines with the Ryzen 7950X nor
> with the Xeon-E2488 reach the performance per core (and no results for
> the Ryzen 9950X have been submitted yet), so it looks like Power 10
> has a really good multi-threading implementation.
> 
> The fact that IBM has not submitted results for Power for SPEC CPU
> 2017 for (Int or FP) Speed or FP Rate results is an admission that
> their numbers there are even less impressive.
> 

That's most likely explanation, but another one is that it is sort of
internal policy no matter what.
IIRC, they didn't publish non-rate scores for POWER7 either, despite
that according to independent measurement at point of introduction
POWER7 single-threaded performance was in the same ballpark with best
Intel offerings and easily ahead of best AMD.

> In any case, certainly for the stuff I do I see no reason why I would
> consider, much less recommend buying a Power machine these days.  My
> guess is that the major reasons for buying pSeries machines these days
> are legacy software and IBM salesmanship.
> 
> - anton

I think, if you are running Oracle DB Enterprise Edition, where
software license per core is the most expensive part then there could
be an economical reason for preferring POWER9 or 10 over Intel or AMD.
But that's just a guess.