9554 and 9654, production modelsWhat are you getting ? model number ? ES ?
Isn't that OS too old to run on Genoa ? Or any modern CPU, server or desktop. Its 13 years old !Motherboard H13SSL-N, Windows 2012 R2 crashes on install - right after loading files and showing Win logo for a second.
Grrrrrrrrrrrrrrr, not happy.
It's supported CPU...Isn't that OS too old to run on Genoa ? Or any modern CPU, server or desktop. Its 13 years old !
Yep, found that and disabled - it stopped crashing there!Try with IOMMU disabled. (It should be in the North Bridge settings.)
Test with OCCT. CPU-Z caused pretty serious coil whine on my 128 thread Epyc ES. I have yet to experience something similar with any other benchmark. Maybe it's just too harsh and unrealistic a benchmark for server multicore testing.Well, 3.55 Ghz all core boost seems BS - or am I doing something wrong?
3.48 is what I get (all core load)Well, 3.55 Ghz all core boost seems BS - or am I doing something wrong? Freqs are not even among cores too, this isn't well cooled area, so might affect things for sure, will retest when they are in proper DC cooled area.
So it only runs 11 memory channels and only runs 2.85 ghz compared to the 9554's 3.48, but it was only $700, compared to the cheapest 9554 was $1100I could return it, but the cheap 9554s are gone. So I guess I keep it.Genoa number 7 up and running. Its a 9534.
I only get 2.7 all core speed, and you said "not well cooled". I plan to get a 360 AIO and then I will hope for more speed, but you seem to be doing well.Well, 3.55 Ghz all core boost seems BS - or am I doing something wrong? Freqs are not even among cores too, this isn't well cooled area, so might affect things for sure, will retest when they are in proper DC cooled area.
Using same stress test from CPU-Z? Perhaps it's lack of air cooling then, which is very plausible given that we put them into tiny assembly office for now3.48 is what I get (all core load)
I only get 2.7 all core speed
9554 is 3.48 and 9654 is 2.7 (needs cooling) Stress test is avx-512 stuff from primegrid, real load.Using same stress test from CPU-Z? Perhaps it's lack of air cooling then, which is very plausible given that we put them into tiny assembly office for now
So is it 2.7 or 3.48, getting confused here, sorry!
---
It's also possible Win 2012 updates lowered down max boost for one reason or another - vanilla install (before multiple waves of patches) was showing CPU at maximum 3.7 ghz, however at that time it was in 32 code mode to get it to work in the first place (win 2012 did not like all 96 cores enabled at start). Problem is that H13SSL-N lacks official support for Win 2012 - so maybe it even runs quicker than it reports, I've seen Win 2012 on supported (!) Icelake show higher clocks than it was capable of.
We've got pairs of those CPUs so will install Win 2019 trial that is supported by chipset drivers and compare side by side in DC cooled environment.
Except with identical coolers, my 9554 runs around 85c, but my 9654 throttles and is at 95c and only runs 2.6-2.7 ghzCooler requirements of 9554 and 9654 are identical. They are both 360 W CPUs. (Or 400 W CPUs if user-configured as such.) And they both reach this power use easily in all-core workloads. (Except for 9554 in rare exceptionally light all-core workloads.)
Cooler requirements of 9554 and 9654 are identical. They are both 360 W CPUs. (Or 400 W CPUs if user-configured as such.)
I wrote about cooler requirements, not about coolers. 360 W are 360 W.¹Except with identical coolers, my 9554 runs around 85c, but my 9654 throttles and is at 95c and only runs 2.6-2.7 ghz
Remember the fineprint:Well, 3.55 Ghz all core boost seems BS - or am I doing something wrong?
"low activity workload" — whatever this means; this may quite possibly exclude SMT usage for example (not to mention vector arithmetic)AMD said:end note (EPYC-021)
All-core boost for AMD EPYC processors is the average frequency of all processor cores running in performance mode while utilizing a low activity workload. Actual achievable all-
core boost will vary based on hardware, software, workloads and other conditions.
I am sure you are right about the way these coolers function. But I know that skillz with custom water cooling (I think) has his 9654 running at like 3.1 ghz all core load. My Turin had the same problem, and an AIO fixed that. So when I get the money, I am getting a 360 AIO for the 9654. I am done with custom water, and I have no data center env. to run a 4U with AC through it.I wrote about cooler requirements, not about coolers. 360 W are 360 W.¹
You have 4U single-tower style coolers, but have them in conventional tower computer cases, not in rackmounted chassis with high air speed. Edit, to be clear, if you strap such a cooler onto a 360 W CPU, you are supposed to blast a huge amount of cubic feet per second of conditioned air through its relatively tiny radiator.
Does anybody around here know if EPYC's (particularly EPYC 9004's) V-f curves are temperature dependent? My guess is that they are, but I don't know for sure. They *should* be, for power efficiency reasons at lower temperature and for computational reliability reasons at elevated temperature. — If the V-f curves are temperature dependent, then the core clocks at 85 °C are likely already reduced, and at 95 °C they most certainly are reduced, compared to better cooled operation.
Actual socket power throttling (as opposed to mere V-f-T related clock reduction) of both 9554 and 9654 happens when T_junction reaches 100 °C (according to cpu-monkey.com, I can't find this data at amd.com right now). I don't have my 9554 running right now, so don't know which of the sensor readings shows T_junction or at least corresponds best with T_junction.
Remember the fineprint:
"low activity workload" — whatever this means; this may quite possibly exclude SMT usage for example (not to mention vector arithmetic)
"and other conditions" — low core temperatures for example?
________
¹) Actually, the chip surface from which these up to 360 W need to be dissipated is larger on the 9654 compared to the 9554. Hence, 9654's 360 W are in fact a little easier to dissipate than 9554's 360W — easier as far as the heat spreader, TIM, and the cooler's cold plate are concerned. But not as far as the cooler's heat pipes, the cooler's radiator, air speed, and inlet air temperature are concerned. And these latter ones are the more important factors in a cooling setup of server CPUs like these.
Was that PrimeGrid too?But I know that skillz with custom water cooling (I think) has his 9654 running at like 3.1 ghz all core load.
don't quote me, but I think so. All of my number are for PGWas that PrimeGrid too?
"low activity workload" — whatever this means
I have not read of anybody using 2012 R2 with ANY CPU, but anything to do with Genoa is welcome here.Are you sure you don't have a H13sssl-nt ? The -N does not have a 10 mbit port. From the manualSo, for the benefit of any other poor sods who want/need to get Windows 2012 R2 work on Genoa 9004 series (and most likely Turin (9005 series) also), using SuperMicro H13SSL-N (BIOS version 3, board revision was v2 - supporting Turin) by default results in crash with error 0x000005C (detailed screen in post above). Should be enough to put most people off, but not me...
Different BIOS settings were tried to no avail, but in the end it seems the following minimal changes make it possible to install -
1) Disable SMT (default was Enabled)
2) In NorthBridge settings -> IOMMU should be Disabled (default was Enabled)
3) Cpu configuration -> SVM mode - disabled (was enabled)
4) CCD control - 2 CCDS (was auto), this is to reduce number of cores active - Windows 2012 seem to be unable to handle more than 64 during install (indicated as spinning dot wheel (under Win logo) forever during boot)
5) In Boot - UEFI BIOS mode should be on (not Legacy) - maybe it will work with Legacy also though, so not 100% sure about it
Vanilla install recognised AMD 9654 by name straight away and that it is 96 cores (with less shown as active). This gets Windows 2012 R2 installed - it's out of support (apart from paid extended one), but old patches are available, takes forever to download them. Once everything is patched CCD control can be put back to Auto, and Windows is happy with full 96 cores on, it's likely SMT can be turned off and MAYBE disabled virtualisation stuff too (not tested as we don't need it).
Things that did not work due to lack of chipset support (boo to SuperMicro - Windows 2012 was still supported when chipset was out!)
1) on board 10 Gbit networking won't work - had to use external card
2) a bunch of hardware bits in Devices is shown as lacking drivers, which they do - nothing appears critically important (to us) at this stage at least...
Still strange annoying thing: Vanilla install shown 96 core as 3.7 ghz (but it had 32 cores enabled only), this is correct top single core speed, however after installing patches max speed seems to have dropped to around 3.1 Ghz, still pending testing in well cooled DC, so will update on that.
Turin next...
It's the last Windows server OS that has NO per core licensing, so when using it with 96 cores it's pretty cost effective - the price of current Windows core packs will be close or higher to cost of MB+CPU+memory. Windows 2012 R2 supports NVMes, can in theory support up to 255 cores too - so savings might be even higher with future processors.I have not read of anybody using 2012 R2 with ANY CPU
CPU-Z is CPU-Z, and your actual planned workload is going to be something else...AMD 9554 (64 cores) shows nice 3.7 Ghz, which is just little shy of 3.75 ghz all core boost, can't complain here
AMD 9654 (96 cores) - just 3 Ghz instead of claimed 3.55 - grrrrr!