Do EPYC Milan CPUs even exist for sale?

lihp

Active Member
Jan 2, 2021
164
43
28
They are not entirely impossible to get, but nowhere near the supply of a year ago.
@Patrick @AlphaG - there is one thing to know: Right now worldwide logistics is really down. Meaning alot of goods are somewhere on hold. Thats in warehouses, ports,... Once everything Corona is less an issue, I am pretty sure we will be flooded with stuff.

Until then: find some Computer shop, who can order at well known distributors like Ingram Micro is always a safe bet. Takes some time, but will work.
 

DRW

New Member
May 1, 2021
7
1
3
Provantage is listing the 7443P hpe upgrade kit as a special order for about the same as the chip is elsewhere. Not sure if that's any quicker for anyone. I think I'll stick with my BLT order, but the ship date has gone from May to June to September.
 

lihp

Active Member
Jan 2, 2021
164
43
28
Here 7443, 72F3 and 73F3 are directly available from stock. Rome are also available for the most part... - like right now.

Thats Germany.

PS - fun fact: 7443P is not available and not even ordered....
 

ca3y6

New Member
Apr 3, 2021
19
2
3
I was also told by a uk supplier I preordered my 7443P with that they are shipping next week, but I am not holding my breath.
 

lihp

Active Member
Jan 2, 2021
164
43
28
I was also told by a uk supplier I preordered my 7443P with that they are shipping next week, but I am not holding my breath.
I now got some weeks testing under my belt with 7443P. Bottom line: impressed.
  • With decent cooling I could get all 24 cores to 3.90+ GHz under prolonged full load
  • 10th in multicore on Passmark: Click Me
  • Single core performance according Passmark makes it the 6th fastest server CPU: CPUbench: Click on Server
  • no hickups so far in burn in test
  • decent price/core
Overall the combination of price:core, single-core, multicore performance, low energy usage (when not in turbo) and amount of cores makes it by far one of the best available cpus as long as you dont need more cores.
 
  • Like
Reactions: nnunn

Hac

New Member
Apr 19, 2021
2
2
3
I ordered 7443P and a motherboard from a European SuperMicro reseller 30.5., at the time they quoted 1-2months for CPU availability.

Today I got a shipping notice, seems I need to really start figuring out ProxMox and planning the setup. Shorter wait than I anticipated, fortunately.
 
  • Like
Reactions: lihp

ca3y6

New Member
Apr 3, 2021
19
2
3
I ordered 7443P and a motherboard from a European SuperMicro reseller 30.5., at the time they quoted 1-2months for CPU availability.

Today I got a shipping notice, seems I need to really start figuring out ProxMox and planning the setup. Shorter wait than I anticipated, fortunately.
Same story here, ordered 24th of March, received a shipping notice today from a uk retailer (novatech). It seems Milan are finally shipping.
 

mirol

New Member
Jan 6, 2018
23
1
3
35
Hi guys I just got email that my 7443p been dispatched today from novatech store in uk will arrive on monday will test it under water I preorder it about month ago but its not on stock on website now so they prob s3nd all preorders and run out ;)So its look some stock is getting in in some places heh.
 

ca3y6

New Member
Apr 3, 2021
19
2
3
I have run some performance test, trying to compress some videos using ffmpeg 4.3.1 (3 times for different files, different resolutions). I am a bit surprised by how it compares to two other machines I have. Of course they have a vastly different number of cores so it's not a fair comparison of the overall throughput if you run multiple compressions in parallel.

ResolutionCodec
i7-6700k​
7302P​
7443P​
960x540h264
6.5x​
14.5x​
6.6x​
960x540h265
4.0x​
6.6x​
3.4x​
1280x720h264
3.8x​
9.8x​
4.4x​
1280x720h265
2.7x​
4.6x​
2.4x​
1920x1080h264
1.6x​
5.3x​
2.9x​
1920x1080h265
1.2x​
2.4x​
1.3x​

At low resolution the i7-6700k holds its ground, which is unsurprising given that there is not much multi-threading going on. What surprises me is that at 1080p, the 7302p vastly outperforms the 7443p even though they have similar base clock (3.0 vs 2.85GHz), and the 7443P has a much higher boost clock (3.3 vs 4.0 GHz), and my 7302P has 2.667GHz RAM vs 3.2GHz for my 7443P. There doesn't seem to be any thermal throttling on the 7443P, it just seems to operate at a lower frequency.

[Edit] See two screenshots below. The 7443P seems to be operating at 2GHz, even though from the noise, the fans seems to be on low power, like if it was trying to save energy:

7302P.PNG7443P.PNG
 
Last edited:
  • Like
Reactions: DRW

lihp

Active Member
Jan 2, 2021
164
43
28
I have run some performance test, trying to compress some videos using ffmpeg 4.3.1 (3 times for different files, different resolutions). I am a bit surprised by how it compares to two other machines I have. ...
Actually I don't know ffmpeg - so nothing I can tell about it.

General
  • Passmark imho gives a rough overview on how performance overall should be.
  • The i7 may come as close as 80-90% to the 7443P in single core performance. The 7302P should be almost 50-60% to the 7443P in single core performance.
  • In multicore it highly depends if ffmpeg takes advantage of highly multithreaded environments. Afaik ffmpeg with its codecs doesnt really put high core machines to their limit. See for example here. Any application taking advantage of high core counts should be along the lines:
    • i7-6700K: 100% (set as baseline) in multicore
    • 7302P: 361% in multicore compared to the i7
    • 7443P: 662% in multicore compared to the i7
Knowing that, let's look at your tests:

ffmpeg
  • Overall ffmpeg is known to be bad at using many cores.
  • The 7302P seems almost ok in relation to the i7 - give or take alot of inefficiencies in its code. Eg. 1920x1080 in h264 comes close to expected values.
  • The 7443P is totally off. Yet since the architecture of 7302P and 7443P are so similar it is expected that the 7443P must outperform the 7302P in all areas (except for some potential weird bugs in ffmpeg none knows about).
  • 3200MHz Memory freq and Per Socket Mem BW of 204.8 GB/s is identical on both AMD CPUs. CPU architecture is also very similar. Meaning both AMD CPUs are highly comparable. In that light the underperformance of the 7443P is baffling.
As a result we need to look at your OS and hardware - thats where the discrepancy must originate (suggested you used ffmpeg in similar settings on each):

OS and hardware
  • For the OS here a good start from Microsoft.
  • Is the board of the 7443P updated with a new BIOS fitting for Milan?
  • Are there any BIOS options you may have missed? (Boost/Turbo freq off, SMP off, ...)
  • Are both AMD Epyc CPUs on ECC DDR4 3200 MHz? Did you test memory (burn in test)? Please check memory performance on both machines and compare.
  • Is the 7443P firmly seated on the CPU socket? Maybe there is a hair between? Something else?
  • Is the 7443P CPU cooler well connected to the CPU? Maybe something wrong between CPU cooler and CPU? One single piece sand can kill everything. Thermal paste ok? Maybe wrong or old thermal paste? I had people where the nice son made fun by putting toothpaste into the thermal paste tube....
  • potential bottlenecks on the 7443P board like storage too slow compared to the 7302P system? Maybe you got a huge RAID, which is still rebuilding on the 7443P on initial start (can take 24+ hours uninterrupted operation depending on storage size)
  • How about the motherboard? PCIe errors? Any conflicts?
  • ...

Bottom line: imho in your case it must be OS and/or hardware or worse some part is broken.
 
  • Like
Reactions: ca3y6

lihp

Active Member
Jan 2, 2021
164
43
28
PS: Which OS are you on anyways? In case of Windows 10 you afaik need Windows 10 Enterprise for Windows server I posted the MS statement above.
 

ca3y6

New Member
Apr 3, 2021
19
2
3
Thanks. on your questions:
  • OS: all 3 are using the most up to date version of Windows Server 2019.
  • The two server boards are H11SSL-i for the 7302P and and H12SSL-i for the 7443P (updated to latest firmware - 2.0 for Milan support).
  • I have not changed any of the BIOS defaults in term of CPU settings, and all CPU features are on "auto" for 7443P rather than "disabled".
  • Both are using ECC, though the 7302P is using slower RAM (2667MHz) than the 7443P.
What makes me think that this is not a fan/socket badly connected is that if I run the same test with two ffmpeg compressions in parallel (so now utilizing all cores on both machines), both are running at their full base clock speed, and the results are in line with what you would expect: 1080p@h264 compressed at 3.16x for 7302P vs 3.7x for 7443P.

It seems that the problem comes from the fact that when only one instance of ffmpeg is running, it is using multiple threads, but not all the cores available, and so 7443P seems to remain in a power saving mode since it is neither a full capacity on a single thread nor it is at full capacity on all cores, whereas the 7302P is closer to its all-core capacity. So the counter-intuitive conclusion is that your individual ffmpeg compressions will run faster if you run two in parallel than if you run only one on a 7443P.

Screenshot of running two ffmpeg in parallel:
7302P2.PNG7443p2.PNG
 

lihp

Active Member
Jan 2, 2021
164
43
28
Did you take a look at those threads on google about ffmpeg and thread switch? Seems you got some potential there :)

Imho 7302 and 7443 should in same settings deliver expected results of 7302 offer app 60% performance of 7443. To me it looked as if the SuMi settings on the H12SSL are quite conservative (when I checked mine). Actually I had 12+ hours testing the board and its settings. Also check of SMP/Hyperthreading is turned on in Windows. Compare performance when its off.