Xeon E7 8894 v4 - what type of Memory for best performance (single core)?
- Thread starter MichalPL
- Start date
-
- Tags
- 8894v4 dl580 gen9 ram ddr4 xeon
Notice: Page may contain affiliate links for which we may earn a small commission through services like Amazon Affiliates or Skimlinks.
That is insane..... Is it the much lower clocks that is the reason?PassMark - AMD Ryzen 9 5950X - Price performance comparison - the result is bit under 46k
Passmark doesn't yet have results for quad E7-8894, but its smaller cousin quad E7-8880 (24 vs 22 cores and 165W vs 150W TPD)
PassMark - [Quad CPU] Intel Xeon E7-8880 v3 @ 2.30GHz - Price performance comparison - 50k
FTFY, and yes.
Im looking at one of these v4 servers for a heavily multithreaded workload.
This will use as many threads as i can get but no one single thread utilizes 100% of its assigned core currently so im thinking i could get away with a whole bunch of lesser cores like the xeon v4.
Also, what impact does removing security mitigations have, i wonder?
OK in multicore performance not that bad (overall rank affected by single core performance drama) - looking on the details. It is like 64 core 3995WX, similar to fastest 64 core Epyc, slower than TR 5995X, but sill should be at least 15-20% faster :/
and yes new passmark fully utilize all 96 cores.
Compare to others (Integer Math and SSE):
4x 8894 v4 (not tweaked yet, problems with boost):
AMD Ryzen Threadripper PRO 5995WX:
[Dual CPU] Intel Xeon Platinum 8380:
AMD Ryzen Threadripper 3970X:
AMD Ryzen 9 5950X:
Intel Core i9-12900K:
and yes new passmark fully utilize all 96 cores.
PassMark Intel vs AMD CPU Benchmarks - High End
PassMark Software - CPU Benchmarks - Over 1 million CPUs and 1,000 models benchmarked and compared in graph form, updated daily!
www.cpubenchmark.net
Compare to others (Integer Math and SSE):
4x 8894 v4 (not tweaked yet, problems with boost):
| Integer Math | 488,554 MOps/Sec |
|---|
| Extended Instructions | 125,986 Million Matrices/Sec |
|---|
AMD Ryzen Threadripper PRO 5995WX:
| Integer Math | 628,857 MOps/Sec |
|---|
| Extended Instructions | 119,620 Million Matrices/Sec |
|---|
[Dual CPU] Intel Xeon Platinum 8380:
| Integer Math | 586,518 MOps/Sec |
|---|
| Extended Instructions | 153,512 Million Matrices/Sec |
|---|
AMD Ryzen Threadripper 3970X:
| Integer Math | 265,668 MOps/Sec |
|---|
| Extended Instructions | 73,828 Million Matrices/Sec |
|---|
AMD Ryzen 9 5950X:
| Integer Math | 192,671 MOps/Sec |
|---|
| Extended Instructions | 40,343 Million Matrices/Sec |
|---|
Intel Core i9-12900K:
| Integer Math | 138,265 MOps/Sec |
|---|
| Extended Instructions | 33,849 Million Matrices/Sec |
|---|
I used numbers from my experience (I "fix" it later and copy paste the numbers like 5995X from passmark - because dont have any) when we build a few 5950X workstations.
+$50 for better AIO plus 3.8GHz DDR4 and 5950x is around 50k passmark points instead of 46k.
Wow it means 4x 8895 v2 on DDR3 is "faster"
It says 57k but I never was able to do more than 51k on dl580 g8.
Might be smart. They are almost for free now, like G8 2.5year ago. The only problem is (for me at least) single core drama. 2.5year ago also was no Ryzen5000, 12900k and Epyc zen3.
Confirmed 8894v4 can do this, I am fighting with 2.9GHz all core limit (was 3.2GHz in 8895v2 - 15 vs 24 cores but also 22 vs 14nm).
Maybe even Oracle 8x 8895v2 DDR3 will be better?
Hmmm, good idea - do you know how to do this?
Did you disable mitigations? And is this score the result after you disabled them?Ok now 8894v4 it's faster than 8895v2 @3.6GHz
2115 points on single core @ 2.88GHz, still need 3.4GHz (or more)
AMD EPYC 7773X - 2410 points :/
AMD EPYC 7773X still a bit faster in "Integer Math" much slower in (not needed in my case) SSE/AVX.
Ok fully tested, the winner in the memory is: 2133MHz 4Rx4 32GB (slightly ~0.3% faster than 2400MHz Rank 2 32GB).
working at 1600MHz CL 12-11-11.
The only current issue is: 2 DIMMs per channel are improving speed a bit, so I should add in the future another 512GB,
in just 2 cartridges populated (1 CPU only connected to the RAM) in the multicore performance was huge (~20%), in all it's less but still visible.
Multicore result is between Threadripper PRO 5995WX and EPYC 7773X now,
single core is 12% lower than EPYC 7773X, and no Idea how to increase it (boosting to 3.4GHz now).
working at 1600MHz CL 12-11-11.
The only current issue is: 2 DIMMs per channel are improving speed a bit, so I should add in the future another 512GB,
in just 2 cartridges populated (1 CPU only connected to the RAM) in the multicore performance was huge (~20%), in all it's less but still visible.
Multicore result is between Threadripper PRO 5995WX and EPYC 7773X now,
single core is 12% lower than EPYC 7773X, and no Idea how to increase it (boosting to 3.4GHz now).
Just read pdf about "Jordan Creek 2" chips (mem buffers for E7 v3/v4) - and I made BIG mistake in number of RAM dimms and memory configuration.
The bad idea:
E7 is quad channel
So 4x 4 =16 dimms
NOT TRUE
good one:
E7 is quad channel but using SMI interface.
SMI almost double the single DIMM speed.
Each "Jordan Creek 2" chip is converting two ddr channels (3 dimm each) to single SMI interface.
So: 4 processors x 4 SMI x2 = 32DIMMs optimal.
In short there is still space for improvment I put just 16dimms here
The bad idea:
E7 is quad channel
So 4x 4 =16 dimms
NOT TRUE
good one:
E7 is quad channel but using SMI interface.
SMI almost double the single DIMM speed.
Each "Jordan Creek 2" chip is converting two ddr channels (3 dimm each) to single SMI interface.
So: 4 processors x 4 SMI x2 = 32DIMMs optimal.
In short there is still space for improvment I put just 16dimms here
So there is a significant improvement in having all of the memory channels populated?
Also how much power does it draw when running this benchmark?
