POWER8 Workstation Launches On Crowdfunding: $4k For Motherboard, $18k For System

[UbuntuRulez]

That benchmark appears to show midrange POWER beating midrange Xeon (albeit simulated) but in turn being beaten by high end Xeon. The accompanying text emphasizes this point, and that high end POWER chips would show comparably higher performance.

If you do the math, an 3.8 GHz high end POWER8 would be 31% faster than the mid end POWER8 used in the benchmark. This means the high end POWER8 is still slower than a high end Intel Xeon.

[bridgman]

Agreed, but we're getting into fairly small differences rather than "crushing".

[gilboa]

In my somewhat dated experience the Power *can* out-perform a Xeon CPU, but it takes specific workloads. (Massively parallel w/ bus-choking I/O on 4S or above).
However, non of this relevant to a single socket workstation board.

I've been looking at non-IBM Power 8 servers (2S) and they do look interesting (both performance and price wise), but the limited memory capacity (and slower DDR3 memory) is an issue.

Never the less, I wish them luck.

- Gilboa

[c117152]

POWER is slower in database benchmarks too. I bet AIX on Intel Xeon runs faster than on POWER8, but IBM is afraid to release benchmarks pitting POWER8 vs Intel Xeon.

http://www.anandtech.com/show/10539/assessing-ibms-power8-part-2-server-applications/8

I dont get how normal, sane people allow IBM to trick them? Have not they checked POWER vs Xeon benchmarks before jumping on the hype train? Or, are they paid by IBM to hype POWER8? How many here, have an affiliation with IBM?

Like others pointed out, it depends on the workloads.

Regarding that benchmark, the conclusions support what I was told: http://www.anandtech.com/show/10539/...pplications/11

As for affiliation, I know a guy who knows a guy who used to work in IBM... Does that count?

[aht0]

12 core CPU system with 96 hardware threads does sound badass though..

[UbuntuRulez]

Like others pointed out, it depends on the workloads.

Yes, it depends on the workloads. However, I have never seen any workload where POWER8 is faster than Intel. Have anyone of you ever seen a benchmark where POWER8 is faster than Intel? Any benchmark?

[gilboa]

Yes, it depends on the workloads. However, I have never seen any workload where POWER8 is faster than Intel. Have anyone of you ever seen a benchmark where POWER8 is faster than Intel? Any benchmark?

Nothing I can share.
I can point out that 8S Xeon scale horribly (DL780G7?), while 8S (and above) power scaled just fine.
Again, nothing relevant to single socket workstation configuration.

- Gilboa

[c117152]

Yes, it depends on the workloads. However, I have never seen any workload where POWER8 is faster than Intel. Have anyone of you ever seen a benchmark where POWER8 is faster than Intel? Any benchmark?

https://www.youtube.com/watch?v=KHHNr5Vqoxs

https://www.hpcwire.com/2015/06/09/ibm-power8-outperforms-x86-on-stac-benchmarks/

From the first page of the google search. Admittedly, specific workloads and applications but it's still real-world loads using real-world software.

[UbuntuRulez]

Nothing I can share.
I can point out that 8S Xeon scale horribly (DL780G7?), while 8S (and above) power scaled just fine.
Again, nothing relevant to single socket workstation configuration.

- Gilboa

So you are basically saying that when benching a single POWER8 vs single x86 we see that x86 is faster as in all the benchmarks I have linked to, but when you scale out to 8 sockets, x86 falls behind? Well, that sounds reasonable. x86 does not scale perfect, Unix scales better. So, x86 is faster on 1-2 sockets but can not tackle large workloads that require 8-sockets or more.

https://www.youtube.com/watch?v=KHHNr5Vqoxs

https://www.hpcwire.com/2015/06/09/ibm-power8-outperforms-x86-on-stac-benchmarks/

From the first page of the google search. Admittedly, specific workloads and applications but it's still real-world loads using real-world software.

The first link, youtube, benches a POWER8 vs an old E5-2697 v2. Why not bench against a E5 v4? That is not fair of IBM to bench against an old x86.

The second link, is about HPC calculations. POWER8 has 96 threads, so it can simulate 96 Monte Carlo paths at once. x86 lags behind in that respect, as x86 does not have that many threads. Also, it says that POWER8 has much higher bandwidth, "192 GB/sec per socket which is 3x higher than x86". Well, in real life RAM bandwidth benchmarks, x86 and POWER8 has roughly the same bandwidth it seems, POWER8 is only around 10% faster. We see that IBM POWER8 reaches

BM S824	4	251,533	253,216	322,399	319,561
IBM S822L	4	252,743	247,314	295,556	305,955
x86 E7 v3	4	230,027	232,092	248,761	251,161

https://www.ibm.com/developerworks/community/wikis/home?lang=en#!/wiki/W51a7ffcf4dfd_4b40_9d82_446ebc23c550/page/Tuning%20a%20full-system%20PowerKVM%20guest

http://www.cs.virginia.edu/stream/by_date/Bandwidth.html