IBM POWER9 Benchmarks vs. Intel Xeon vs. AMD EPYC Performance On Debian Linux

csbin

Senior member
Feb 4, 2013
816
8
136
#1
https://www.phoronix.com/scan.php?page=article&item=power9-epyc-xeon&num=1

The Raptor Talos II test system was configured with the aforementioned dual eight core IBM POWER9 CPUs yielding 64 threads, 256GB of system memory, a 500GB HDD (what they had available, but I/O testing thus isn't a focus for today's tests), and an AMD Radeon Pro WX 7100 graphics card. The POWER9 CPUs clock up to 3.80GHz.

The system was loaded with Debian Testing and using the Linux 4.16 PPC64LE kernel. The Linux 4.15 kernel and newer currently support the Talos II. Our Xeon/EPYC testing was done with Debian Testing as well using its GCC 7.3 stock compiler, EXT4 file-system, and upgrading to the Linux 4.16 kernel.

The other x86 server/workstation systems I had locally for comparison to these initial POWER9 benchmark numbers were:

2 x Intel Xeon Gold 6138 - The Tyan S7106 1U barebones with two Xeon Gold 6138 processors yielding a combined 40 cores / 80 threads, 96GB of RAM, and Samsung 850 EVO 256GB. Xeon Gold 6138 processors have a 2.0GHz base frequency and 3.7GHz turbo frequency.

AMD EPYC 7551 - The Gigabyte MZ31-AR0 with an AMD EPYC 7551, which is 32 cores / 64 threads at a 2.0GHz base clock frequency, 2.55GHz all core boost speed, and 3.0GHz maximum boost clock speed. This system had 32GB of memory and NVIDIA GeForce GTX 970 graphics.

AMD EPYC 7601 - The TYAN B8026T70AE24HR with an AMD EPYC 7601, which is 32 cores / 64 threads with a 2.2GHz base frequency, 2.7GHz all core boost frequency, and 3.2GHz maximum boost clock speed. This system had 128GB of DDR4 and a 280GB Intel 900p Optane SSD.

Those were the systems available for this initial round of testing against the 64-thread Talos II Workstation while all systems were on Debian Testing with the Linux 4.16 kernel. This is just the raw performance results due to not being able to deliver performance-per-Watt metrics at this time from the Talos II.
 

Zstream

Diamond Member
Oct 24, 2005
3,179
8
116
#2
https://www.phoronix.com/scan.php?page=article&item=power9-epyc-xeon&num=1

The Raptor Talos II test system was configured with the aforementioned dual eight core IBM POWER9 CPUs yielding 64 threads, 256GB of system memory, a 500GB HDD (what they had available, but I/O testing thus isn't a focus for today's tests), and an AMD Radeon Pro WX 7100 graphics card. The POWER9 CPUs clock up to 3.80GHz.

The system was loaded with Debian Testing and using the Linux 4.16 PPC64LE kernel. The Linux 4.15 kernel and newer currently support the Talos II. Our Xeon/EPYC testing was done with Debian Testing as well using its GCC 7.3 stock compiler, EXT4 file-system, and upgrading to the Linux 4.16 kernel.

The other x86 server/workstation systems I had locally for comparison to these initial POWER9 benchmark numbers were:

2 x Intel Xeon Gold 6138 - The Tyan S7106 1U barebones with two Xeon Gold 6138 processors yielding a combined 40 cores / 80 threads, 96GB of RAM, and Samsung 850 EVO 256GB. Xeon Gold 6138 processors have a 2.0GHz base frequency and 3.7GHz turbo frequency.

AMD EPYC 7551 - The Gigabyte MZ31-AR0 with an AMD EPYC 7551, which is 32 cores / 64 threads at a 2.0GHz base clock frequency, 2.55GHz all core boost speed, and 3.0GHz maximum boost clock speed. This system had 32GB of memory and NVIDIA GeForce GTX 970 graphics.

AMD EPYC 7601 - The TYAN B8026T70AE24HR with an AMD EPYC 7601, which is 32 cores / 64 threads with a 2.2GHz base frequency, 2.7GHz all core boost frequency, and 3.2GHz maximum boost clock speed. This system had 128GB of DDR4 and a 280GB Intel 900p Optane SSD.

Those were the systems available for this initial round of testing against the 64-thread Talos II Workstation while all systems were on Debian Testing with the Linux 4.16 kernel. This is just the raw performance results due to not being able to deliver performance-per-Watt metrics at this time from the Talos II.
That Epyc system sure loves clocks. It’s soet of amazing what a mere few hundred mghz can achieve.
 

CatMerc

Golden Member
Jul 16, 2016
1,113
53
106
#3
That Epyc system sure loves clocks. It’s soet of amazing what a mere few hundred mghz can achieve.
More likely it's because it has 4x the RAM and an Optane SSD.
 

Topweasel

Diamond Member
Oct 19, 2000
4,893
498
136
#4
I don't get the review. I do get that the IBM systems would be great for use in certain processes and are fast enough for workloads that can't easily be ported over to x86, but realistically that system considering it price, the amount of memory, isn't really that competitive. I mean it got mostly thrashed in the tests. Probably entering good enough territory in terms of how much closer it is than previous systems but it really didn't stand apart in any test and was near the bottom on most of it. But at the end of the test the reviewer suggests that it is good option without displaying that in the tests. Heck I could be completely wrong and with just a little work on software and the Power9 would shame AMD and Intel. Just nothing in that test even came close to displaying it.
 
Oct 12, 2014
115
28
101
#5
I don't get the review. I do get that the IBM systems would be great for use in certain processes and are fast enough for workloads that can't easily be ported over to x86, but realistically that system considering it price, the amount of memory, isn't really that competitive. I mean it got mostly thrashed in the tests. Probably entering good enough territory in terms of how much closer it is than previous systems but it really didn't stand apart in any test and was near the bottom on most of it. But at the end of the test the reviewer suggests that it is good option without displaying that in the tests. Heck I could be completely wrong and with just a little work on software and the Power9 would shame AMD and Intel. Just nothing in that test even came close to displaying it.
Considering it's only 16 P9 cores, against 32 or more x86 cores, often running code streams with mature assembly optimizations for x86, I think it does rather well. And it's not particularly expensive, either; the 8-core chip under test has a list price under US$600. The Xeon Gold Phoronix compares it to has a list of US$2612, and the Epyc 7551 seems to have pricing around US$3500 or a little higher. (I don't see a public list price table for Epyc, but most places seem to be selling it for US$3500-4000, and Anandtech says "above $3400.")

If we're comparing Power9 against CPUs with pricing in the mid 4 digits, the more natural place to do it is with the 18-core or 22-core Power9 SKUs, which cost $1375 and $2575 respectively.
 

Topweasel

Diamond Member
Oct 19, 2000
4,893
498
136
#6
Considering it's only 16 P9 cores, against 32 or more x86 cores, often running code streams with mature assembly optimizations for x86, I think it does rather well. And it's not particularly expensive, either; the 8-core chip under test has a list price under US$600. The Xeon Gold Phoronix compares it to has a list of US$2612, and the Epyc 7551 seems to have pricing around US$3500 or a little higher. (I don't see a public list price table for Epyc, but most places seem to be selling it for US$3500-4000, and Anandtech says "above $3400.")

If we're comparing Power9 against CPUs with pricing in the mid 4 digits, the more natural place to do it is with the 18-core or 22-core Power9 SKUs, which cost $1375 and $2575 respectively.
I couldn't find any pricing when I searched best I could find was referencing $5-6k and I couldn't really tell if it was a the top or bottom of the pricing window. If their that cheap then great. But the point still stands about the review/test. We can assume that the jobs are that much slower because it lacks the core count of the other systems tested. But all of the systems there have wildly different specs with one EPYC having an 8th of the memory. So then at the end after it basically tail ended every test, the writer says it's great CPU. I understand testing with what you have on hand, it's just poor optics to test a piece of hardware with a test suite you choose have it preform poorly in those tests in comparison and then call it great.
 
Oct 12, 2014
115
28
101
#7
I couldn't find any pricing when I searched best I could find was referencing $5-6k and I couldn't really tell if it was a the top or bottom of the pricing window. If their that cheap then great. But the point still stands about the review/test. We can assume that the jobs are that much slower because it lacks the core count of the other systems tested. But all of the systems there have wildly different specs with one EPYC having an 8th of the memory. So then at the end after it basically tail ended every test, the writer says it's great CPU. I understand testing with what you have on hand, it's just poor optics to test a piece of hardware with a test suite you choose have it preform poorly in those tests in comparison and then call it great.
Pricing is at https://www.raptorcs.com/ - has 4core, 8core, 18core, and 22core SKUs.

I agree the optics suck, but Phoronix has a long and proud history with really weird hardware comparisons. I don't get it either. It's still not fair to say P9 is non-competitive because low-end P9 got compared by a Linux news site against systems with 2-2.5x the core count and far higher CPU price.
 

Markfw

CPU Moderator, VC&G Moderator, Elite Member
Super Moderator
May 16, 2002
18,000
1,809
136
#8
I would like to see a comparison based on price. If I were a data center manager (and I was very close to that in the job I retired from), I would have a budget. Then I would compare performance/reliability/power usage on different platforms, and get the best. So, like $30k servers and then 60k servers, and the like. Based on my threadripper experience, I think EPYC would be competitive on all fronts, but I have no real stats to prove that.
 
Last edited:

Topweasel

Diamond Member
Oct 19, 2000
4,893
498
136
#9
Pricing is at https://www.raptorcs.com/ - has 4core, 8core, 18core, and 22core SKUs.

I agree the optics suck, but Phoronix has a long and proud history with really weird hardware comparisons. I don't get it either. It's still not fair to say P9 is non-competitive because low-end P9 got compared by a Linux news site against systems with 2-2.5x the core count and far higher CPU price.
Cool thanks for the link. Yeah my point or at least what I was trying to get to wasn't so much meant to be a statement of Power9. But how little you see of it's strength, ones that the writer proclaims, in that test.
 

Zstream

Diamond Member
Oct 24, 2005
3,179
8
116
#10
Cool thanks for the link. Yeah my point or at least what I was trying to get to wasn't so much meant to be a statement of Power9. But how little you see of it's strength, ones that the writer proclaims, in that test.
Yeah, but power consumption is crazy high on power9 CPUs.
 

Atari2600

Senior member
Nov 22, 2016
803
273
106
#11
Based on my threadripper experience, I think EPYC would be competitive on all froms, but I have no real stats to prove that.
Apparently, in terms of some cryptos, EYPC absolutely kicks Threadrippers arse (scaling far beyond what you'd expect).

Dunno why, but it is what it is.

[Now, obviously, I'm not making this statement for use on cryptos, rather more, workloads can see EYPC outperform TR by an unexpected scale.]
 

eek2121

Senior member
Aug 2, 2005
294
6
116
#12
Apparently, in terms of some cryptos, EYPC absolutely kicks Threadrippers arse (scaling far beyond what you'd expect).

Dunno why, but it is what it is.

[Now, obviously, I'm not making this statement for use on cryptos, rather more, workloads can see EYPC outperform TR by an unexpected scale.]
same die, the only way it would beat threadripper is more cores. Threadripper walks all over anything involving AES. It gets more than twice a 20x Xeon E5-366- v3 with much faster memory according to AIDA64. EPYC is based off the same Zen arch, just lower MHz.
 

ehume

Golden Member
Nov 6, 2009
1,382
17
91
#13
I think the unsaid point of the review is the entirely made-in-the-USA system was price-competitive and performance-competitive with systems built partially or wholly with Asian-produced components. This is an important point for Defense Dept related hardware, and gives the manufacturer a selling point with a purchaser with very deep pockets.
 

Markfw

CPU Moderator, VC&G Moderator, Elite Member
Super Moderator
May 16, 2002
18,000
1,809
136
#14
I think the unsaid point of the review is the entirely made-in-the-USA system was price-competitive and performance-competitive with systems built partially or wholly with Asian-produced components. This is an important point for Defense Dept related hardware, and gives the manufacturer a selling point with a purchaser with very deep pockets.
OK, I miss that one. Which is totally US ?
 

beginner99

Diamond Member
Jun 2, 2009
4,155
218
126
#15
More likely it's because it has 4x the RAM and an Optane SSD.
And the power9 is the only one with a HDD. No wonder it looses very clearly in certain tests.
 

ehume

Golden Member
Nov 6, 2009
1,382
17
91
#16
OK, I miss that one. Which is totally US ?
Just look at the specs. The first-reviewed machine with the PPC's. I will have to beg forgiveness: the RAM and the HD's are passive, might be sourced outside the US.
 

Melo89

Junior Member
Jun 11, 2018
1
0
1
#17
More likely it's because it has 4x the RAM and an Optane SSD.
I don't think the RAM and SSD matter for the LBM fluid dynamics test? The epyc 7551 and 7601 have the same l3 cache and nearly the same clock rates. Why would epyc 7601 beat 7551 by almost a factor of 2x, assuming simulation fits in RAM? I'm not sure I understand this result (see OpenMP LBM fluid dynamics test).
 


ASK THE COMMUNITY

TRENDING THREADS