Do EPYC Milan CPUs even exist for sale?

jpmomo

Active Member
Aug 12, 2018
348
118
43
@Keith Myers Just thought I'd let you know I just build a H12SSL + 7313P combo and it consistently turbos to 3.725 GHz under almost all workloads. Including heavy load like prime95 small. I'm surprised because it's above the official turbo of 3.7 GHz (maybe they needed more significant digits in their marketing)? LOL

P.S. I still really badly need a 7443P in north america.. can anyone help me find one!?!

one option is to just purchase a whole server if that works for you. I think someone else on this thread did the same but with a lenovo server. I also purchased a couple of servers from dell that had the milan cpus. It obviously costs more and might not work if you are trying to build your own server. Provantage has an hpe server that comes with the 7443p and has 27 in stock. unfortunately they raised the price on the system from 2700 just a week ago to 3100 now! shopblt had a few of the same servers for a bit less but they are now sold out (went from 6 to 0 in 1 day!)

 

Spartus

Active Member
Mar 28, 2012
223
36
28
Toronto, Canada
Thanks for the advice buying whole servers. I might just do it. at 2700 I would have stomached it I think.

I have been on shopblt backorder for a few weeks, that's where i got the 7313p.

Thanks for the advice, it gave me some new options.
 

Keith Myers

Member
Oct 10, 2020
95
17
8
I changed my Determinism control to Power from Performance on my 7443P and now am getting all the cores to hold up at max boost clocks. The AMD tuning docs recommend Performance and are incorrect for my workload.

Fought getting the 7402P configured because my BIOS must have got corrupted because Determinism disappeared from the menus. Even flashing a new BIOS or an older BIOS didn't get the parameter to appear. Had to pull the battery and clear the BIOS to fix the corruption. Have Power configured now for that host also. Much happier that the clocks are staying up closer to max boost.
 

Spartus

Active Member
Mar 28, 2012
223
36
28
Toronto, Canada
Yeah I got tricked by that at first. It would seem intuitive that "performance" gives the best performance, but its actually meant for consistent performance. so it throttles the turbo. Power gives consistent power control allowing turbo to hit the max clocks but staying at or under the cTDP setting you configure.

(or I'm out to lunch on my explanation, but regardless, use "power")
 

mirrormax

Active Member
Apr 10, 2020
141
62
28
Yeah I got tricked by that at first. It would seem intuitive that "performance" gives the best performance, but its actually meant for consistent performance. so it throttles the turbo. Power gives consistent power control allowing turbo to hit the max clocks but staying at or under the cTDP setting you configure.

(or I'm out to lunch on my explanation, but regardless, use "power")
Yea I think only some really niche use cases would want to use performance mode, so each node performs the same no matter variation in silicon. but it is def confusing since performance intuitively sounds faster, and it's no described well in the bios from what i remember
 

Keith Myers

Member
Oct 10, 2020
95
17
8
Well I was using the AMD Epyc 7002 tuning guide previously and Performance was suggested for all the HPC configurations.
But looking at the new AMD Epyc 7003 tuning guide, they have updated it and explain the differences between the two determinism profiles much more clearly.
• Power: Ensure maximum performance levels for each CPU in a
large population of identically configured CPUs by throttling CPUs
only when they reach the same cTDP

• Performance: Ensure consistent performance levels across a large
population of identically configured CPUs by throttling some CPUs
to operate at a lower power level

So cranking up cTDP and PPL to the max limit for each cpu will gain you running at max or close to max boost clocks under all-core loading.
 

Keith Myers

Member
Oct 10, 2020
95
17
8
No. I can only reach nominal 3.6Ghz on the 7443P with my current cooling. I am changing to a better block on Thursday when my Optimus block arrives. With that block I can get the 7402P to within 5 Mhz of max boost clock. I am hoping the better block will allow similar with the 7443P.
 

lihp

Active Member
Jan 2, 2021
179
44
28
No. I can only reach nominal 3.6Ghz on the 7443P with my current cooling. I am changing to a better block on Thursday when my Optimus block arrives. With that block I can get the 7402P to within 5 Mhz of max boost clock. I am hoping the better block will allow similar with the 7443P.
From my experience:
  1. "Normal load tests" - tests for up to 1 h with cold environment and great cooling. Basically high load with 30-50% average load and burst loads: there I can achieve up to 3.97 - 4.02 on almost all cores at up to 70°C core temperature.
  2. "Hammer loads" - tests with 70% - 95% average load with prolonged 100% spikes for 3+ hours straight (thats what I am doing right now): up to 3.5 - 3.6 GHz at 75°C - HT enabled.
So for that extreme load I gotta consider different cooling if I want to push further ;). On a sidenote: besides the CPU load, I am also running block-IO up to 30GB/s write and 10GB/s read (NVME RAID) and 2+GB/s on large block RAID arrays - meaning the system is overall under real stress right now. TBH: at that amount of load I am quite content...

Unbenannt.jpg
 

Keith Myers

Member
Oct 10, 2020
95
17
8
I also wonder if the architectural differences between Rome and Milan are affecting the max sustained work load clocks.

Using the same BIOS settings and memory and cooling except for the cpu blocks.

My loading is 93% constant BOINC loading at 53° C. Clocks between 3.583-3.602 Ghz

But for a like for like example, my 5950X using Zen 3 cores runs much hotter than my 3950X using Zen 2 cores. At almost the exact same Vcores.
Something to do with the improved floating point pipelines and increased L3 caches.
 
Last edited:

ectoplasmosis

Member
Jul 28, 2021
98
36
18
To add an extra data point:

I have several 7443P systems on Asrock ROMED8-2T boards, with Noctua NH-U9 TR4-SP3 air coolers running the stock Noctua 92mm fans.

When running a prime95 torture test, which uses AVX instructions and is a worst-case 100% load, the CPU package power remains at a reported 205W, the temperatures of all cores settles at 70°C, and the all-core boost settles at ~3.2GHz.

When running a real-world sustained parallel workload, such as video encoding/rendering using ffmpeg/Cinema4D, all-core boost fluctuates between 3.2-3.6Ghz, with an average of ~3.5GHz, again with sustained steady temperatures of 70°C.

Pretty impressive for a 24-core 205W air-cooled CPU.
 

Keith Myers

Member
Oct 10, 2020
95
17
8
Thanks for those datapoints.

I just wish I could figure out the magic recipe I have running on the EPYCD8 and 7402P and apply that to the ROMED8-2T and 7443P.

Same workloads but very different end results with respect to max boost clock achieved sustainably.
 

ectoplasmosis

Member
Jul 28, 2021
98
36
18
Thanks for those datapoints.

I just wish I could figure out the magic recipe I have running on the EPYCD8 and 7402P and apply that to the ROMED8-2T and 7443P.

Same workloads but very different end results with respect to max boost clock achieved sustainably.
It seems like the 7443P’s sustained all-core boost under 100% load is limited by its 200W TDP to ~3.5GHz, less if using AVX.

I don’t think you’re doing anything wrong, just hitting the power envelope ceiling.
 

Keith Myers

Member
Oct 10, 2020
95
17
8
But I hit the same sustained 200W TDP on the 7402P with the same exact workload and able to maintain the max boost clock of 3.35Ghz on the 7402P within 5 Mhz. No AVX load just floating point calcs.

Based on the logic of same workload and power budget, the 7443P should be able to do the same. But it doesn't.

A team colleague says the difference is the absolute clock on both processors. World of difference between 3.35Ghz and 4.0Ghz.

And stop whining and enjoy the fact I am one of the few North American DIY builders that have been able to lay hands on a 7443P.

He is still waiting on his pre-order of the 7443P.
 

ectoplasmosis

Member
Jul 28, 2021
98
36
18
But I hit the same sustained 200W TDP on the 7402P with the same exact workload and able to maintain the max boost clock of 3.35Ghz on the 7402P within 5 Mhz. No AVX load just floating point calcs.

Based on the logic of same workload and power budget, the 7443P should be able to do the same. But it doesn't.
I've got six self-built systems running 7443Ps and they all behave the same.

The point I'm trying to make is that the 7443P is capped at a maximum package power usage of 200W, and in the case of your specific workload, it hits 200W at 3.6GHz all-core. Under different workload scenarios, you'll find it hitting different all-core boost frequencies, but still hitting that same 200W package power ceiling.

The fact that your 7402P can hit 99% of its boost clock at all-core load is irrelevant; it's still bumping up against its 200W TDP, just the 'max boost' frequency is different between the two CPUs, and the 7402P can only ever boost as high as 3.35GHz regardless of package power load.

There is no guarantee of a specific all-core boost frequency with either product, rather each CPU will boost as high as it can in all situations, within its TDP and max boost ceilings.

Your CPU is behaving as it should, and your colleague is correct in the sense that you should relish the 4.0GHz max boost ceiling during low thread count tasks ;)
 

Keith Myers

Member
Oct 10, 2020
95
17
8
Your CPU is behaving as it should, and your colleague is correct in the sense that you should relish the 4.0GHz max boost ceiling during low thread count tasks
Agreed, it's just that I will never see that 4Ghz under my load as it is never light. Only light loading is when I haven't started my normal workloads and am simply benchmarking for giggles.
 

ectoplasmosis

Member
Jul 28, 2021
98
36
18
24x Zen3 cores plus IO die simply need more than 200W to hit over ~3.5GHz all-core under full load.

It's a great chip for the price, especially compared to the 'frequency-optimised' 74F3, which will give you a few more MHz at 240W for over twice the price...
 

Keith Myers

Member
Oct 10, 2020
95
17
8
I agree, it's a great "bang for the buck" product. Finally the 7443P has caught up and passed my 7402P in production.
Shouldn't complain. Just enjoy the latest and greatest.
 

DRW

New Member
May 1, 2021
18
6
3
So I have a 7443P finally. After my last post, I cancelled that order with BLT and ordered the HPE upgrade kit instead. I got that last week. I had some issues getting it to boot, but it turned out to be a faulty ram module. Swapped it out with another I had (same part number), and it booted fine. Waiting on a satadom as the one I got was 7-pin power, and I guess supermicro only supports 8-pin.
 
  • Like
Reactions: lihp