So anyone got a Ryzen 5000 yet? DC benchmarks? ;)

Page 3 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

StefanR5R

Elite Member
Dec 10, 2016
5,459
7,718
136
F@H and Rosetta have dramatically large workunit variability.
These need to be tested with a fixed workunit. But nobody has come up with such a test yet.

(There is FahBench, but AFAIU it is somewhat disconnected from F@H "production".)

It's possible that a fully reproducible PrimeGrid PPS-DIV benchmark will be posted here sometime soon.
 

StefanR5R

Elite Member
Dec 10, 2016
5,459
7,718
136
PS,
don't get me wrong: Please run the projects which you want to run, by all means. All what I am saying is that meaningful performance metrics for F@H and Rosetta are not easy to come by. (Déjà vu. There was once the idea to collect perf/W data for Rosetta.)
 

Endgame124

Senior member
Feb 11, 2008
954
669
136
PS,
don't get me wrong: Please run the projects which you want to run, by all means. All what I am saying is that meaningful performance metrics for F@H and Rosetta are not easy to come by. (Déjà vu. There was once the idea to collect perf/W data for Rosetta.)
I understand - mostly I’ve only run 8 projects - seti, WCG OP, F@H, climate, Milky Way, Einstein, asteroids, and Rosetta. Climate isn’t great because it doesn’t always have work, especially for windows. Of the others, the only ones I’ve run more than 1 thread on a Pi are WCG OP, F@H, and Rosetta. The “numbers” projects I have very little interest for because they just seem to run to find “neat” things like large primes - I’m happy to setup a test if it would be valuable to someone, but they aren’t projects im likely to run outside of a pentathlon.
 

StefanR5R

Elite Member
Dec 10, 2016
5,459
7,718
136
I'm harboring the thought to put offline tests together for other projects besides PrimeGrid for a long while now. Who knows, perhaps I'll overcome my inertia some day.

I went through with this with PrimeGrid's LLR based subprojects because these put a stringent requirement on the user to find out at least useful, if not optimum, application configurations for each computer × each subproject. At other projects, you generally just run the application and don't tune anything anymore.

Still, for questions like how is bang-for-the-buck of 3950X vs. 3900X or 3950X vs. 5950X if one is mostly interested in World Community Grid, or what PPT limit setting could be the best compromise at TN-Grid, or should I disable turbo on my Intel processor at SiDock, or what precise mix of concurrent MIP and OPN work will give me highest MIP throughput, … reproducible ( = offline) tests in more projects would be good to have.
 
  • Like
Reactions: Assimilator1

biodoc

Diamond Member
Dec 29, 2005
6,257
2,238
136
The top computer at Wanless (linux only) is a 5950X. The run times on the majority of tasks are quite consistent. I'm running my 3950X on Wanless too so I decided to compare run times and PPD on 40 completed and validated tasks. A few short run time tasks appear occasionally so I eliminated those from the comparison.

5950X:
Average run time: 2,076.25 sec (STDEV = 6.7 sec)
Average points per tasks: 28.95 (STDEV = 0.33 points)
PPD: 38554.74

3950X @ 3.9 GHz: PPT@105 watts
Average run time: 2,293.61 sec (STDEV = 3.13 sec)
Average points per tasks: 28.74 (STDEV = 0.48 points)
PPD: 34647.97

The 5950X showed a 9.5% decrease in task run time and a 11.3% increase in PPD.

I guess we were expecting ~10-20% increase in IPC with the Zen 3 when compared to Zen 2, correct?
 
  • Like
Reactions: Assimilator1

Endgame124

Senior member
Feb 11, 2008
954
669
136
The top computer at Wanless (linux only) is a 5950X. The run times on the majority of tasks are quite consistent. I'm running my 3950X on Wanless too so I decided to compare run times and PPD on 40 completed and validated tasks. A few short run time tasks appear occasionally so I eliminated those from the comparison.

5950X:
Average run time: 2,076.25 sec (STDEV = 6.7 sec)
Average points per tasks: 28.95 (STDEV = 0.33 points)
PPD: 38554.74

3950X @ 3.9 GHz: PPT@105 watts
Average run time: 2,293.61 sec (STDEV = 3.13 sec)
Average points per tasks: 28.74 (STDEV = 0.48 points)
PPD: 34647.97

The 5950X showed a 9.5% decrease in task run time and a 11.3% increase in PPD.

I guess we were expecting ~10-20% increase in IPC with the Zen 3 when compared to Zen 2, correct?
Zen 3 on average is supposed to be 19% faster than zen 2, including the clock speed increase, so strictly speaking IPC that looks right
 

Endgame124

Senior member
Feb 11, 2008
954
669
136
Updated the bios on my crosshair hero viii to the latest beta bios. Haven’t had a chance to try higher ram clocks, but PBO pushed the all core clock speed to 4550mhz.
 

Endgame124

Senior member
Feb 11, 2008
954
669
136
first batch of information from my 5950.

It runs without fuss at 3733 ram 1866 infinity fabric clock and 1866 U clock. I can run 4000 ram 2000 IF and uclk, but windows is barely stable at stock voltage - if I add load, it crashes, so I probably need to tweak voltage.

with PBO enabled and set to motherboard, and curve optimizer set all cores at -20, I get 4.65 ghz all cores. Whole system Power draw varies between 301w and 306w when running 32 threads of WCG OP. Average points per day over 3 days was 240,682. This is slightly lower than it should be, as I rebooted multiple times playing with fabric speeds and curve optimizer.

With Eco mode enabled, power draw is locked solid at 145w and clock is pretty solid at 3775. No points yet for eco mode.
 

Endgame124

Senior member
Feb 11, 2008
954
669
136
first day results show eco mode as 89% of the points but half the power draw. Almost seems too good to be true, will need a few more days to get an average.

other comparisons:

the 5950x on eco mode yields roughly 15.5times the points per day than my A10-7970K at the same power budget. It yields 3.27x the points per day at slightly less power than my 2700X at all cores at 4.4ghz. It yields almost exactly 36 times the points per day of a raspberry pi 4, while using around 35 times the power, making the 5950x slightly more power efficient than the pi4.
 

Endgame124

Senior member
Feb 11, 2008
954
669
136
I’m partial to the water cooled pi cluster myself:
System-Running-And-Booted-Up-1536x864.jpg
 

Endgame124

Senior member
Feb 11, 2008
954
669
136
Wow. What a clean setup! That 120mm radiator seems like overkill to me though. ;)
Yeah, I think a 120mm is totally overkill for only 8 pis. I probably would have loaded up a 16 port or 24 port switch with Pis to really make use of the system, and at that point would have considered Power over Ethernet for a little more efficiency power wise. When it comes to the water cooling, I would have likely routed the whole thing differently too - I'd put the res at the top for easy filling, added quick disconnects to each side so I would only need to drain half for any maintenance, and then found a way to run them all in parallel instead of serial as the last pi in serial probably would be getting a little warm.

The build log is here (https://www.the-diy-life.com/building-a-water-cooled-raspberry-pi-4-cluster/ ) if you're interested in looking at it - maybe someone should recruit him for distributed computing since he said he doesn't know what to do with the whole thing.
 

Endgame124

Senior member
Feb 11, 2008
954
669
136
3 day average for the 5950 on eco mode was very consistent. 390, 395, and 395 work units. Average point per day were 223,348. I’m going to switch the host to F@H for the contest, but haven’t decided what to do for the next 6 days. If anyone has any wishlist items, let me know
 
  • Like
Reactions: Assimilator1

Assimilator1

Elite Member
Nov 4, 1999
24,120
507
126
Milkway, yes on the CPU ;), and if you could, on my benchmark thread post the averaged time of at least 6 227.5x credit WUs, ta :)
 
  • Like
Reactions: Endgame124

Assimilator1

Elite Member
Nov 4, 1999
24,120
507
126
I was going to suggest Rosetta too (especially seeing as it's a good CPU project), but other than letting the RAC build up (which you don't have time for atm), I don't know how to benchmark it.
 

Endgame124

Senior member
Feb 11, 2008
954
669
136
I was going to suggest Rosetta too (especially seeing as it's a good CPU project), but other than letting the RAC build up (which you don't have time for atm), I don't know how to benchmark it.
Agreed on Rosetta - I’ll do that after the F@H race. I’ll also be able to compare with raspberry pis, my 2700x, and a i3-6100. might be the last thing I use the 2700X for, as I intend to upgrade that PC to a 5950 too.

I just setup Milky Way to run, and it pulled down a bunch of GPU jobs for the GeForce 710. Bleh, each of those are going to take an hour to process.
 

Assimilator1

Elite Member
Nov 4, 1999
24,120
507
126
I'll have those times too! lol
And yea you'll need to disable the app in the website preferences to avoid GPU WUs, or just suspend it :p
 

StefanR5R

Elite Member
Dec 10, 2016
5,459
7,718
136
Not directly DC related, but Ryzen 5000 related:

I found this link in the comment section beneath AnandTech's i7-10700 review:
https://cdn.mos.cms.futurecdn.net/i9W8M8HgGaTqRs4bVn5A39-1920-80.png

I haven't taken the time yet to look up from which article this graph came from. Until I do so, I am left wondering: Have CPU reviews at Tom's Hardware become a whole lot more useful (to DCers, specifically) than AnandTech's? (It's long ago that I last browsed Tom's.) — OK, it says merely "estimated task energy", not measured task energy. But still.
 

StefanR5R

Elite Member
Dec 10, 2016
5,459
7,718
136
Re: CPU reviews which are useful to DCers, specifically:

In AnandTech's 5980HS review of today, the 15 W TDP config performs almost the same as the 35 W TDP config in quite a lot of tests. In some of these tests, it is either obvious from the bar graph or a little bit less obvious from the description in the article that these are very short running tests — hence not reflecting performance in 24/7 duty. Some other of these "15 W ≈ 35 W" tests were long running. I presume that these test were only lightly threaded, such that the 35 W config was not power limited in this test — hence not reflecting performance in DC which can almost always use all cores fully.

[I am not saying that such tests are useless. They are in fact quite useful, especially to laptop buyers. I am merely saying that these two classes of tests are not indicative of performance in Distributed Computing, obviously. In this laptop processor review, this rather trivial fact sticks out prominently, but it may easily be overlooked in desktop or workstation CPU reviews.]
 
  • Like
Reactions: Assimilator1

Markfw

Moderator Emeritus, Elite Member
May 16, 2002
25,483
14,434
136
This is a little OT, but maybe the 5000 series, after todays Q4 results will have better supply ?? Lets hope so.
 

Endgame124

Senior member
Feb 11, 2008
954
669
136
This is a little OT, but maybe the 5000 series, after todays Q4 results will have better supply ?? Lets hope so.
Wccftech said great supply of 5600 and 5800 in q1. Limited supply of 5900 / 5950. Not the most reliable site though
 
  • Like
Reactions: Ken g6