NAS Build: Xeon vs. EPYC 3000 vs. Ryzen

aivxtla

New Member
Apr 19, 2020
10
5
3
Also unlike your approach looking at AIDA64, I believe STH and that other review both took into account the power draw for the entire system for both the 3251 and 2141, ie BMC, RAM etc. Anandtech I believe had a similar approach with their D-1540 review if I’m not mistaken.
 
Last edited:

111alan

Member
Mar 11, 2019
98
29
18
Haerbing Institution of Technology
Also unlike your approach looking at AIDA64, I believe STH and that other review both took into account the power draw for the entire system for both the 3251 and 2141, ie BMC, RAM etc. Anandtech I believe had a similar approach with their D-1540 review if I’m not mistaken.
Reviewers usually use a Watt meter or the current sensing functions on a more complexed PDU for overall power tests. I brought up the clamp meter earlier to see how much 12V power the CPU VRM directly drains from PSU. If everything else other than CPU takes 30-40w at least(with main PSU efficiency already accounted for), there's no way that EPYC platform can idle at 29w.

The two board X11SDV-8C-TLN2F and M11SDV-8C-LN4F doesn't seems to differ from each other component wise, except EPYC one has way better heatsink and more VRM phases(to handle more power draw?). BTW I don't think that simple heatsink on Xeon can handle 100w without some large helicopter fan though...
 
Last edited:

aivxtla

New Member
Apr 19, 2020
10
5
3
Well I'm not sure what else is at play here to account for the massive idle differences other than the differing CPU architectures. I agree its a bit too high, even for the 2123 which is 4C/8T. Maybe due to Intel QAT? I'd assume not at idle though.


1608589062158.png
 
Last edited:

111alan

Member
Mar 11, 2019
98
29
18
Haerbing Institution of Technology
Well I'm not sure what else is at play here to account for the massive idle differences other than the differing CPU architectures. I agree its a bit too high, even for the 2123 which is 4C/8T. Maybe due to Intel QAT? I'd assume not at idle though.


View attachment 16824
QAT is also avaliable in Xeon Scalable. A 4-core 2.xGHz CPU at more than 50w idle and more than 100w max is unthinkable. Doubt they're testing with some lab PSU that doesn't even meet 80plus white std, or have some other things in the system to drain power.

BTW 2123IT only has 1 AVX3 unit, which was made by combining the existing two AVX2 units, so there won't be any extra power used for AVX3 instructions.
 
  • Like
Reactions: aivxtla

Evan

Well-Known Member
Jan 6, 2016
3,287
567
113
I don’t understand at all why D-2100 as a small SOC uses so much power especially at idle. It’s not like it’s even high frequency either.

D-1541 systems I have idle with 128gb just below 30w at the wall. I would have expected the D-2100 would be similar but shocked when the reviews showed otherwise by a large margin.

In the full Xeon range we didn’t see an increase like this from broadwell to skylake just to make it even more confusing.
 

aivxtla

New Member
Apr 19, 2020
10
5
3
I finally got a chance to turnoff the 3251 system and connect the Killa-A-Watt. After bootup the entire system effectively shows 30 Watts idle (with 2x 8GB DDR4 2666 RDIMMs, an Intel X710-T2L dual port 10Gbe adapter with both ports actively taking traffic albeit low load conditions, 2x Crucial MX100 256GB SSDs (RAID 1), 1x Samsung 960 PRO 512GB (Inactive), and 3x 8500 RPM 40mm fans at ~1,500-1,800 RPM.)
 
Last edited:

diskdiddler

Member
Mar 3, 2017
84
7
8
43
I recall wanting a new server and really getting excited about the prospect of an Epyc 3000 system for great performance but also low power.

I ended up with Denverton, because no one out there, seemed to want to put more than 4 SATA ports on an Epyc 3 series processor.

Anyhow, the Epyc 3004 series (Zen 4) is rumoured for next year now, disapointingly, it apparently will have a baseline wattage of 65. Hopefully that information is incorrect. I'd be happy with a 35w, 3ghz model with 8 cores myself.

Anyhow, thought people may be interested. I'm guessing part numbers will be something like "Epyc 3254" for example.
 

EffrafaxOfWug

Radioactive Member
Feb 12, 2015
1,395
501
113
I think it's more than just a rumour thankfully - there was a roadmap slide I saw a few weeks ago showing the new embedded processors. The image in the article below seems to indicate they'll sit in a 65-120W power envelope but I suspect there'll still be low-power options available.


Like you I was also waiting for an Epyc 3000 NAS platform (ended up building a Ryzen server instead); lack of SATA ports was an issue, and then the lack of any updates to the 3000 silicon in light of the substantial Zen 2 and Zen 3 improvements made them even less compelling. Let's hope the OEMs will be nicer to these little chips the next time around, and hopefully also spur Intel on to improve its Atom offerings...
 
Last edited:
  • Like
Reactions: TXAG26

NateS

Active Member
Apr 19, 2021
124
64
28
Sacramento, CA, US
I think the reason SATA3 ports are becoming less common on high end server platforms is that there's only so many HSIO pins on the CPU, and they'd rather use those for PCIe (each lane of which can drive a whole lot of SATA ports via an AIC), and a lot of servers today don't use SATA anyway. Personally, I don't see it as a huge disadvantage given how many PCIe lanes are available and how cheap HBAs are these days. Do you have a particular use case for needing a lot of SATA ports on the motherboard, rather than on an AIC?
 

lunadesign

Active Member
Aug 7, 2013
218
26
28
Do not use AMD CPUs for i/o intensive or idle power sensitive workloads. Using IOmeter(and FIO in linux) I observed a 20-40% IOPS decrease for high-end NVMe drives like PM1725a and Pblaze5, and it doesn't scale nearly as well when you add more drives. Although the difference will be smaller in not so CPU bound scenarios such as sequential read, ...
Focusing only on the performance aspect of this, has anyone else observed results like this?
 

NateS

Active Member
Apr 19, 2021
124
64
28
Sacramento, CA, US
I don’t understand at all why D-2100 as a small SOC uses so much power especially at idle. It’s not like it’s even high frequency either.

D-1541 systems I have idle with 128gb just below 30w at the wall. I would have expected the D-2100 would be similar but shocked when the reviews showed otherwise by a large margin.

In the full Xeon range we didn’t see an increase like this from broadwell to skylake just to make it even more confusing.
One thing to keep in mind is that the D-2100 series has an entire Lewisburg PCH embedded in it, which contains 4x 10gbe and has a 15W TDP all by itself. That may not account for all of the increased power consumption over the previous generation, but it's certainly a big chunk of it.
 

111alan

Member
Mar 11, 2019
98
29
18
Haerbing Institution of Technology
Focusing only on the performance aspect of this, has anyone else observed results like this?
Well you should search if there are any people who secretly did not get this problem. There is a convenient way for this: the Crystal Disk Mark happen to include the 4KB random 16T32Q test by default. A 400-500K IOPS translates to about 1.6-2GB bandwidth in that test. A high end enterprise drive like PM1725a or Pblaze5 series should have 820K+ IOPS, that is at least 3.2GB bandwidth in that test. Just browse the internet to see if any Zen2 platform got that value. GL.

BTW Zen3 solved it. But not for Gen4 SSDs(about 1M IOPS bottleneck for a 1.5M IOPS drive).
 
Last edited:

lunadesign

Active Member
Aug 7, 2013
218
26
28
Yikes. I'm still learning a lot about NVMe SSDs and wish I had seen this before I started a new storage server build using an Epyc 7262. I was lured in by the # of PCIe Gen 4 lanes, reasonable power consumption and chip availability.

On the bright side, I should still be way ahead of my current all-SATA setup on an Xeon E5-1650 v2. :)