Zen 2 for Distributed Computing: Any interest?

[pututu]

@VietOZ , awesome hardware that you got there. I"m still running old xeon boxes (v2 and v3).

 

[Assimilator1]

asking the Stealth cooler to cool what we're doing is a bit much. Honestly, I don't think even the popular Hyper 212 can cool these 7nm chips for crunching and that's a 4 heat pipes cooler. The Hyper 212 can handle any V3/V4 xeon, but won't survive with these new Ryzen. Unless you undervolt the crap out of it and that'll cause clock stressing. My 3600 non X is running at 4ghz with 1.32v vcore all core and cooled by a 280mm AIO (Corsair) still hits low 70s with avx WUs. 70s are high for me .. it's just personal preferences.... so I lowered my cpu to 3.9ghz kept the temp under control. I don't like my cpu or gpu goes over 70C.

Not really, the stock cooler is supposed to be able to keep the CPU (@ stock settings) cool enough that it can run any app (except avx I'd imagine) without thermal throttling. Bear in mind I'm running with the side of the case off, & I'm using high grade thermal paste too.
Although I think SETI uses avx, I don't think F@H does, does it?

 

[biodoc]

Although I think SETI uses avx, I don't think F@H does, does it?

Yes, I believe F@H uses avx according to this thread and others at the folding forum.

 

[Assimilator1]

Ah ok, that changes things somewhat. Although I still would've thought it should keep it cool enough for it to at least not throttle........

 

[Markfw]

@TennesseeTony , I think you wanted to see WCG scores with a lot of cores. Well, I have 170 on Rosetta, and 495 on WCG, so tomorrows stats should show that. And I may be bringing 64 more on-line if the power does not give out.

 

[VietOZ]

@pututu Thanks bro! Cost me arms and legs, but I love hardware. Glad that I only have a couple of hobbies and the other one (wood working) is not as expensive.
@Markfw looking to see your numbers on WCG. I'm very interested in those Rome Epyc.

 

[Markfw]

@pututu Thanks bro! Cost me arms and legs, but I love hardware. Glad that I only have a couple of hobbies and the other one (wood working) is not as expensive.
@Markfw looking to see your numbers on WCG. I'm very interested in those Rome Epyc.

Look here: https://www.worldcommunitygrid.org/...atus=4&projectId=-1&sortBy=sentTime&pageNum=1

Remember, it runs 128 at a time, and only runs 2 ghz fully loaded.

 

[plonk420]

hrm, not showing up here (and i'm logged into wcg)

 

[Markfw]

hrm, not showing up here (and i'm logged into wcg)

How about this ?

 

[StefanR5R]

Look here: https://www.worldcommunitygrid.org/...atus=4&projectId=-1&sortBy=sentTime&pageNum=1

hrm, not showing up here (and i'm logged into wcg)

It seems WCG doesn't show host stats except to the host's owner. But 3rd party stats sites do (if the owner enabled the corresponding export option in the privacy settings). Here is Mark's device ID 6313713:

https://stats.free-dc.org/host/bwcg/6313713
https://www.boincstats.com/stats/15/host/detail/6313713​

 

[StefanR5R]

I should downsize my DC gear in terms of heat output, and for that I need to look into upgrade options with an efficiency level that currently only the 64-core SKUs of EPYC can provide, and maybe the 48-core SKUs come close to. Anything else would merely be a sidegrade with insubstantially improved efficiency, it seems.

To follow up on this: As reported elsewhere, last month I went with an 32-core SKU after all, mainly for cost reasons, secondarily for per-socket cooling requirements. Thirdly, I still am wondering about memory bandwidth per core — contrasting 2-CCD per memory controller vs. 1-CCD per memory controller —, but I guess the former is still fine for virtually all DC projects.

32-core Rome provides circa double the power efficiency of my previously most efficient computers, and I'm OK with that. I failed to decrease the total heat output of my DC gear though; I only added to it.

:-)

I am looking forward to the Threadripper 3000 announcements which should happen this month, but one is for certain: At stock they won't be as energy-efficient as EPYC 7002. Whether or not BIOS settings will allow to drive their energy use down to EPYC levels will remain to be seen.

I lost interest in TR3K as soon as specs of it and its platform came out. For the time being, I have no use for factory overclocked processors, nor for all the power-hungry I/O and oversized mainboards.

 

[Assimilator1]

What did the 32core Rome replace? Great efficiency gain anyway!

I get lost with all these different CPUs lol, so a quick bit of Googling shows both Epyc 1 & TR 3000 are Zen2 CPUs but on different platforms it seems.
Epyc aimed at servers, so have lower overall clock speeds (for efficiency presumably). The EPYC 7502P & TR 3970X are both 32c/64t CPUs, with the latter a ~1.2 GHz advantage.
Here in UK atm, for these 2 CPUs 3970X ~£1770, EPYC 7502P ~£2400.
Are the sTRX4 mbrds a lot more expensive than the SP3s as well as more power hungry? (I only looked so far ).

 

[ao_ika_red]

I lost interest in TR3K as soon as specs of it and its platform came out. For the time being, I have no use for factory overclocked processors, nor for all the power-hungry I/O and oversized mainboards.

From watching video below, I immediately know TR is operated beyond its efficiency curve. I mean, TRX40 has huge array of VRM compared this Epyc board.

 

[Endgame124]

From watching video below, I immediately know TR is operated beyond its efficiency curve. I mean, TRX40 has huge array of VRM compared this Epyc board.

The draw for the threadripper is for those who need cores and clock speed without ECC and other server features, and who may not be able to spread load across multiple hosts.

At this point, though, only a few months away from the next iteration, might as well see what the new chips bring.

 

[Assimilator1]

How do the mbrd prices compare?

 

[Endgame124]

How do the mbrd prices compare?

Similar - $400 to $850 depending on features like built in 10gbit LAN

 

[StefanR5R]

What did the 32core Rome replace? Great efficiency gain anyway!

I still have all my other computers, and of course use them all in Pentathlon. My most efficient ones were dual E5-2696v4 (2x 22c/44t Broadwell-EP, Intel's first 14 nm server processor generation).

Are the sTRX4 mbrds a lot more expensive than the SP3s as well as more power hungry? (I only looked so far ).

Idle power consumption of my PCIe v3 based dual socket Rome computer (here) is the same as as idle consumption of sTRX40 computers (there). As a small difference, these sTRX40 test systems had a GTX 1080 in them which idles at 8 W or less, whereas my computer is headless (but then again, has 4 times as many DDR4 channels).

Current UK retail prices of mainboards seem to be £ 440...690 for SP3 and £ 380...950 for sTRX40.

 

[TennesseeTony]

And somehow, with typical German precision, Stefan could take even your grand-mother's 486DX2, and get more points than I can with a 3950X.... Much respect, friend!

 

[StefanR5R]

A pro pos 486DX2, my own main desktop at home is still based on what feels like an immediate successor to 486DX2 nowadays, a Haswell Xeon E3. It sits on a Supermicro entry-level workstation mainboard. I am vaguely interested in updating this to an equivalently rock-solid but less power-using modern system. In the DC context, it will be functioning more as a console, less as a worker, just like this E3 v3. Therefore, I am interested in very low idle power consumption (or at least very low platform power consumption), something which AMD currently does not have on offer at desktops or servers. Maybe the Epyc Embedded devices approach this use case, but they are not updated to Zen 2 yet, nor do they have an IGP.

 

[Assimilator1]

I finally got the bolt through adapter kit for my 12yr old Thermalright Ultima 90 yesterday , swapped out the crappy Stealth cooler for it & unsurprisingly got a 15C drop in temps for LHC (and ~+50 MHz clock interestingly) , the same avx WUs. In Rosetta the drop is even more dramatic at 20C! Although strangely still a lower clock speed (about 3.725 GHz vs 3.8 GHz for LHC), I don't get that, does anyone else?
My next step is to set the PPT back to default & see if it'll hold reasonable temp's.

Then I need to check the 'Power reporting deviation' to see if that was also responsible for the high temps with the stock cooler & PPT setting.
Has anyone else hear read about that? "Ryzen Burnout? AMD Board Power Cheats May Shorten CPU Lifespan" - Tom's
The CPU lifespan thing is apparently over-egged, but it could certainly (in part) explain the temperature problem I was having, we'll see.

Btw, is it true that the recent batch of Ryzens overclock much better?

 

[Markfw]

I finally got the bolt through adapter kit for my 12yr old Thermalright Ultima 90 yesterday , swapped out the crappy Stealth cooler for it & unsurprisingly got a 15C drop in temps for LHC (and ~+50 MHz clock interestingly) , the same avx WUs. In Rosetta the drop is even more dramatic at 20C! Although strangely still a lower clock speed (about 3.725 GHz vs 3.8 GHz for LHC), I don't get that, does anyone else?
My next step is to set the PPT back to default & see if it'll hold reasonable temp's.

Then I need to check the 'Power reporting deviation' to see if that was also responsible for the high temps with the stock cooler & PPT setting.
Has anyone else hear read about that? "Ryzen Burnout? AMD Board Power Cheats May Shorten CPU Lifespan" - Tom's
The CPU lifespan thing is apparently over-egged, but it could certainly (in part) explain the temperature problem I was having, we'll see.

Btw, is it true that the recent batch of Ryzens overclock much better?

Read Anandtech's Ian Cuttress's response to that. Bottom line ? Don't worry about your CPU dying. As for the lower clocks, no idea.

 

[Assimilator1]

Yea The Stilt said a similar thing over at the HWiNFO forum, and he was the source of the info in the 1st place .
I don't see Ian's response in their (other than a quote by someone).

Now if I could only find out who/where it was that someone was telling me my Ultima 90 wouldn't be up to the job , I thought it was in this thread.......

 

[Markfw]

Yea The Stilt said a similar thing over at the HWiNFO forum, and he was the source of the info in the 1st place .
I don't see Ian's response in their (other than a quote by someone).

Now if I could only find out who/where it was that someone was telling me my Ultima 90 wouldn't be up to the job , I thought it was in this thread.......

This is the article by Ian

Electromigration: Why AMD Ryzen Current Boosting Won't Kill Your CPU

www.anandtech.com

 

[Markfw]

Another 1950x sold... And another EPYC 7742 ES chip is just waiting for 128 gig ram ! It will be here Tuesday.

 

[VirtualLarry]

That's a thing of tech beauty, Mark.

I've got another Ryzen R5 3600 in, a "fresh batch" one, and I've acquired another MSI Tomahawk MAX B450 board (at MSRP no less!), and another Asus ROG STRIX B450-F Gaming ATX board (also @ MSRP!). Deciding whether to do some new builds, some gaming PCs to sell, or to add to my DC fleet.

I'm not so happy with the (lack of) cooling in my Rosewill Magnetar cases (no top venting, no top rad. mount).

I also picked up a couple more of those CoolerMaster MasterLiquid 240 240mm AIO LC kits. ($54.99, seemed reasonable.)

I think that I sent one of those to you, Mark, ages ago, with that Ryzen 2700X I sold you, but I'm sure that you've gotten rid of that by now, now that you're going all EPYC on us.