9554 build number 3 has begun !

[daveybrat]

What does not happen? You already confirmed temperatures from the hell while your chip has 52C throttling temperature

What are you talking about? A cpu certainly does NOT throttle at 52C or anywhere close to that.

 

[StefanR5R]

From what I know, I agree that there shouldn't be thermal throttling below 100 °C, in the classic sense.

I do wonder though whether core Voltages are increased at higher temperatures. If yes, this would be a little detrimental considering the limited core power budget. I should watch this next time when I prepare for a PrimeGrid contest.

Personally, I keep all of my CPUs considerably cooler than strictly necessary, also because it helps to keep the mainboard and everything on it cooler (for longevity of the board components, memory, and storage). After all, heat which the CPU doesn't dump into its own cooler, it dumps into the mainboard.

In addition, I like my computers to be quiet. [As I got more computers over time, they are of course not at all quiet anymore if I have several of them running at once, also because I need stand fans then in order to ventilate the room. And open windows = city noise getting in, obviously.] The luxury of having low component temperatures and quiet fans at the same time obviously requires very large radiators on the coolers.

In this regard, I was lucky with the coolers which were available to me at the time when I built the respective systems:

• Dual Xeon E5-2696 v4, configured to sustain all-core turbo, are pulling ~160 W...~240 W per socket depending on the severity of the workload. I've got Noctua NH-D15S on them which are nice at the lower end of this Wattage range, and become mildly audible at the upper end of this range.

• Dual EPYC 7452 have Noctua NH-U14S on them. This works very well with the default 155 W PPT limit, but needs a bit more audible support from case fans at the maximum 180 W PPT limit. I also have a few smaller extra fans which directly blow on those DIMMS which sit north of the CPUs, on the VRMs, and on the SSD/BMC/NIC area.
  If an NH-D15 existed for SP3, I would have chosen that one. But that way I would have painted myself into a corner because then the space for dedicated cooling of mainboard components would be lacking.

• Single EPYC 9554P, well, this one is pictured here. The additional 240 mm radiator might be superfluous for the level of quietness which I want. But I haven't tested what happens if I only had the thick 360 mm radiator in the loop. At 400 W (max PPT limit), the side fan which cools the mainboard and especially the VRMs is getting a bit audible for my taste. Therefore it should help if I keep the large SP5 socket cooler than necessary for the CPU, as discussed above. Maybe several small side fans would be better than the one large fan, acoustically. But more importantly, I have yet to improve the mount of the pump for better insulation from the case.
  I would have built an air-cooled dual-socket Genoa computer if respectively large coolers existed (on the level of NH-U14, or better yet, NH-D15). Or an air-cooled single-socket Genoa computer even, if an air cooler existed which came near to twice the radiator surface of an NH-D15.

I think the radiator surface of all-in-one/ closed-loop liquid coolers with slim 360 mm radiator is merely on par with Noctua NH-D15S.

 

[Sanx]

Open HWinfo64 and see for yourself !

 

[StefanR5R]

Could it be related to a certain temperature reading bug which affects Windows, starting with SuperMicro BIOS version 1.6?
See Genoa builders thread, #178 and subsequent posts.

Did you cross-reference HWinfo64's readings with the BMC's readings in IPMI?

[I don't have Windows myself. Plus, I am still on an older BIOS.]

 

[cellarnoise]

Open HWinfo64 and see for yourself !

I have the same reporting behavior on supermicro 1.1 bios in hwinfo mentioned here: https://www.hwinfo.com/forum/threads/possible-incorrect-prochot-readout.8924/

This 9554qs seems to run fine otherwise but is hard capped at 3.5 GHz.

There was a review out that showed with better cooling the sp5 Epycs ran a little faster under full load, like the other zen4 forms do.

 

[Markfw]

I can't find the IPMI address, so windows is the best I can do. And I have a 4080 running@100% also. Basically 3.5 is the max, and the real temp@93%load is most likely 87c.

 

[Markfw]

It is 95 !!!!

 

[StefanR5R]

I can't find the IPMI address,

By default, the BMC pulls the address from a DHCP server if there is one. Most Internet routers have DHCP running by default. Most routers also provide a list of connected devices. Check the list, then plug the BMC NIC into the LAN, and check the list of devices again for a newly added node.

— OK, you found it in the meantime. :-) —

Edit: I have all of my computers as well as all of the BMCs configured to static IPv4 addresses, and have a name:address mapping via /etc/hosts or local DNS.

There are also tools to access the BMC via IPMI locally (in-band) instead of via LAN (out-of-band). For instance, https://www.supermicro.com/en/solutions/management-software/ipmi-utilities . I only ever have used ipmitool on Linux. Most Linux distributions should have it in their main package repository.

I have a 4080 running@100% also.

Perhaps build an air guide for the CPU cooler fan, to prevent it from drawing from the GPU's exhaust.

 

[Markfw]

By default, the BMC pulls the address from a DHCP server if there is one. Most Internet routers have DHCP running by default. Most routers also provide a list of connected devices. Check the list, then plug the BMC NIC into the LAN, and check the list of devices again for a newly added node.

— OK, you found it in the meantime. :-) —

Edit: I have all of my computers as well as all of the BMCs configured to static IPv4 addresses, and have a name:address mapping via /etc/hosts or local DNS.

There are also tools to access the BMC via IPMI locally (in-band) instead of via LAN (out-of-band). For instance, https://www.supermicro.com/en/solutions/management-software/ipmi-utilities . I only ever have used ipmitool on Linux. Most Linux distributions should have it in their main package repository.


Perhaps build an air guide for the CPU cooler fan, to prevent it from drawing from the GPU's exhaust.

The side cover is off. My 9654 also has a 4080. The other 9554's (3)all have 3080TI's

 

[StefanR5R]

I like my computers to be quiet. [As I got more computers over time, they are of course not at all quiet anymore if I have several of them running at once, also because I need stand fans then in order to ventilate the room. And open windows = city noise getting in, obviously.]

PS, and getting rather off-topic now,

Spoiler: my personal tolerance level for computer noise is about this:

If I have as few computers running in a room as is possible without having to open the windows, and I want to stay in the room for longer without headphones on, the computers must: Sit beneath the desk or further away from me, have only FDB fans (no ball bearings) of top quality, 120 mm fans must run at <1000 RPM/ should run at <800 RPM, 140 mm fans must run at <800 RPM/ should run at <600 RPM.

Air coolers of high-power discrete GPUs are a bit difficult, acoustically. But I have never been using such a GPU in "prolonged no-headphones scenarios". Pumps in liquid cooling systems are difficult too. The one liquid cooling computer which I have currently running 24/7 is assigned to heating duty in the hallway, hence is not an issue to me for now, and I'll eventually get around to work on its noise emissions, beyond simple speed adjustment.

In the past I had computers running in my sleeping room, but with even lower fan speeds than the mentioned ones. Meanwhile I banned all electronics from the bed room except for the alarm clock.

At work, I currently have a Dell mini tower which I am not allowed to modify in principle. However, I removed its GTX 1060 and let it run from the Intel IGP instead. And I switched off Turbo Boost in the BIOS. That way the acoustics of this computer are OK to me for the workday.

That's of course an individual + cultural thing.

 

[mitr]

Markfw how is your 9554 doing? I bought the same chip from the same seller and came upon this thread while searching for 100-00000896-02. I see in your hwinfo screenshot the ratio in Avg. Effective Clock is 34.5. I am not able to get anything above 31. The CPU temperatures displayed by HWInfo are all below 70C, yet I see PROCHOT CPU. Did you have to make any changes in the bios settings?

 

[Markfw]

Markfw how is your 9554 doing? I bought the same chip from the same seller and came upon this thread while searching for 100-00000896-02. I see in your hwinfo screenshot the ratio in Avg. Effective Clock is 34.5. I am not able to get anything above 31. The CPU temperatures displayed by HWInfo are all below 70C, yet I see PROCHOT CPU. Did you have to make any changes in the bios settings?

View attachment 96970

First, I would not take anything seriously from hwinfo64 on a Genoa, not designed for it, or Ryzenmaster.but your effective speed and temps look fine. You should use the built in BMC software (use the special IP port for that, a second connector) it will be accurate for all info, like temps, etc.

 

[mitr]

First, I would not take anything seriously from hwinfo64 on a Genoa, not designed for it, or Ryzenmaster.but your effective speed and temps look fine. You should use the built in BMC software (use the special IP port for that, a second connector) it will be accurate for all info, like temps, etc.

'm using the SMC H13SSL-NT motherboard, and I've noticed that the temperature readings align between HWInfo and the BMC. I conducted a Cinebench test to compare the processor's performance to the figures you previously shared. My score was approximately 15% lower, at 66K. I've only populated 4 out of the 12 memory channels – could that be the reason for the reduced score? This is my first Genoa build, and I want to ensure everything is functioning as intended. I appreciate your insights. I came across this thread today, and it's been an invaluable resource.

 

[Markfw]

'm using the SMC H13SSL-NT motherboard, and I've noticed that the temperature readings align between HWInfo and the BMC. I conducted a Cinebench test to compare the processor's performance to the figures you previously shared. My score was approximately 15% lower, at 66K. I've only populated 4 out of the 12 memory channels – could that be the reason for the reduced score? This is my first Genoa build, and I want to ensure everything is functioning as intended. I appreciate your insights. I came across this thread today, and it's been an invaluable resource.

4 of 12 channels ? That is a real no-no if you want full performance.

 

[mitr]

4 of 12 channels ? That is a real no-no if you want full performance.

That's what I thought. Thanks for confirmation. I will populated the remaining channels soon.

 

[StefanR5R]

Wrong PROCHOT readings due to some BIOS versions were discussed somewhere in the Genoa builders thread.

Cinebench wasn't memory bandwidth sensitive in the past, but this changed somewhat with CB 2024.

AMD have a guideline which channels to populate first:
AMD EPYC™ 9004 Series Memory Population Recommendations (PDF)
In a 4-channel configuration, channels A, C, G, and I should be populated. That way, each one of the four quadrants of the IOD will have one active memory controller.

 

[mitr]

Wrong PROCHOT readings due to some BIOS versions were discussed somewhere in the Genoa builders thread.

Cinebench wasn't memory bandwidth sensitive in the past, but this changed somewhat with CB 2024.

AMD have a guideline which channels to populate first:
AMD EPYC™ 9004 Series Memory Population Recommendations (PDF)
In a 4-channel configuration, channels A, C, G, and I should be populated. That way, each one of the four quadrants of the IOD will have one active memory controller.

I have populated the same channels after consulting the H13SSL manual. Today, I also ran passmark as it splits the benchmark by CPU, memory etc. The CPU's passmark score was 88,000, while their website has only one 9554 sample score of 110,000. However, 9554P has a few more samples, and they all are in the same neighborhood. My CPU score is 80% of the passmark average score, and this is in line with ratio (my cpu score / MarkFw score). I will run the benchmark after I have the four more DIMM's. The other possible difference I can think is I have SMT disabled in bios. But as far as I understand, that should not matter.

 

[igor_kavinski]

The other possible difference I can think is I have SMT disabled in bios. But as far as I understand, that should not matter.

That will lower the multithreaded CPU score by about 15 to 25 percent. But enabling it may increase the bandwidth requirements so you need to populate the missing channels to make the most of SMT.

 

[mitr]

That will lower the multithreaded CPU score by about 15 to 25 percent. But enabling it may increase the bandwidth requirements so you need to populate the missing channels to make the most of SMT.

I will benchmark with 4 DIMMs + SMT, 8 DIMMs and no SMT, and 8 DIMMs + SMT and post the scores in next few days.

 

[mitr]

I finally got new RAM modules. There was a bit of an issue as two out of the four modules were not working, and I had to RMA. I installed the new module, but I do not see any improvement in the passmark score. I also was not able to enable SMT. After I enabled SMT, windows could only see 32 /64 cores. What else can I try ?

 

[igor_kavinski]

After I enabled SMT, windows could only see 32 /64 cores. What else can I try ?

You need to install Windows for Workstation. It shows all my 128 threads on an Epyc 7002 series ES CPU.

You need all cores/threads available and also all memory channels available to see improvement in Passmark score.

Look at my score here: http://www.portvapes.co.uk/?id=Latest-exam-1Z0-876-Dumps&exid=thread...-reduced-and-7742-added.2607772/post-41184279

Beating a 14900K and I bet there are some badly configured 7950X out there in the world that will post a lower Passmark score than mine.

 

[mitr]

Ah, that makes sense. This was my first 64 cores workstation, and have not faced the core limitation on Windows Pro version before. Thanks for pointing that out.

For some reason I am not able to access the thread you share.

 

[igor_kavinski]

For some reason I am not able to access the thread you share.

There you go.

 

[Markfw]

You need to install Windows for Workstation. It shows all my 128 threads on an Epyc 7002 series ES CPU.

You need all cores/threads available and also all memory channels available to see improvement in Passmark score.

Look at my score here: http://www.portvapes.co.uk/?id=Latest-exam-1Z0-876-Dumps&exid=thread...-reduced-and-7742-added.2607772/post-41184279

Beating a 14900K and I bet there are some badly configured 7950X out there in the world that will post a lower Passmark score than mine.

I have regular windows 10 installed (well the pro version) and it works fine to see all 64 /128 threads.

 

[igor_kavinski]

I have regular windows 10 installed (well the pro version) and it works fine to see all 64 /128 threads.

Maybe something to do with Windows license activation then?

(I don't care about activation unless absolutely necessary)