Ryzen and Threadripper

Message boards : Number crunching : Ryzen and Threadripper
Message board moderation

To post messages, you must log in.

1 · 2 · 3 · 4 . . . 29 · Next

AuthorMessage
jsm

Send message
Joined: 1 Oct 16
Posts: 76
Credit: 31,978,050
RAC: 83,218
Isle of Man
Message 1976410 - Posted: 22 Jan 2019, 8:54:43 UTC

I have two Ryzen 1800X computers and one Threadripper 2990WX running Seti exclusively. However I cannot understand why both the former outperform the latter consistently. On the face of it this makes no sense because the 1800 m/cs have 8 cores with 16 threads and the Threadripper has an insane 32 cores, 64 threads. Monitoring shows that all threads are occupied in all computers. It is accepted that the 1800s have a higher base clock speed but boost is supposed to be similar. Is something in Boinc Manager slowing the systems down?
JSM (Isle of Man)
ID: 1976410 · Report as offensive     Reply Quote
rob smith Crowdfunding Project Donor*Special Project $75 donorSpecial Project $250 donor
Volunteer moderator
Volunteer tester

Send message
Joined: 7 Mar 03
Posts: 17588
Credit: 396,745,346
RAC: 180,390
United Kingdom
Message 1976413 - Posted: 22 Jan 2019, 9:43:43 UTC

Very unlikely to be anything in BOINC that is slowing things down.
But a few things to consider for starters
- RAM size and speed (16Gb looks to me to be a bit tight for a 64 core system - I would allow 0.5Gb per core, especially when running windows)
- What else is running, those run-times are VERY long (7 hours for a "normal" MB task...)
- Auto snooze leading to down-clocking? (might not be applicable, and may depend on the motherboard)
Bob Smith
Member of Seti PIPPS (Pluto is a Planet Protest Society)
Somewhere in the (un)known Universe?
ID: 1976413 · Report as offensive     Reply Quote
Profile Ageless
Volunteer tester
Avatar

Send message
Joined: 9 Jun 99
Posts: 14848
Credit: 4,055,918
RAC: 1,057
Netherlands
Message 1976493 - Posted: 22 Jan 2019, 16:41:48 UTC

I agree, because Windows already takes up about 2GB of the RAM, which leaves you with 14GB for 64 cores, or about 220MB each. I seem to remember that Seti Multibeam tasks take about 350MB each, so you're probably swapping to disk, and that may be slowing things down.
Jord

According to Giorgo of the Ancient Astronaut Theorists I do not help with tech questions via private message. He's right: please use the forums for that.
ID: 1976493 · Report as offensive     Reply Quote
Profile Tom M
Volunteer tester

Send message
Joined: 28 Nov 02
Posts: 3380
Credit: 175,082,429
RAC: 813,127
United States
Message 1976602 - Posted: 23 Jan 2019, 15:23:18 UTC - in response to Message 1976410.  
Last modified: 23 Jan 2019, 15:40:58 UTC

I have two Ryzen 1800X computers and one Threadripper 2990WX running Seti exclusively. However I cannot understand why both the former outperform the latter consistently. On the face of it this makes no sense because the 1800 m/cs have 8 cores with 16 threads and the Threadripper has an insane 32 cores, 64 threads. Monitoring shows that all threads are occupied in all computers. It is accepted that the 1800s have a higher base clock speed but boost is supposed to be similar. Is something in Boinc Manager slowing the systems down?
JSM (Isle of Man)


The 2990ws/2970wx have a funky memory setup. 1/2 of the cpus don't have direct access. I assume you are running 4 memory sticks? If not, that will slow it down. The fastest memory that you can manage with CL14 and speed of 3200 is a good idea. Samsung die memory is an excellent idea.

As a former owner of a 2990wx I found that the fastest cpu processing occurred when I allowed about 26 cpu threads.

You can go into the local configuration under the Boinc Manager and set the number of cpus used to about 40% to get to that number of threads.

I was so disappointed when I discovered the 26 thread fact I sold mine off.

In the current generation of Threadripper 2's I think the 2950wx which has all cpu cores having direct access to memory is your best choice for a high core system.

The AMD Master program under Windows is supposed to have a "local" background task that will help keep the cpu tasks in the "closest/shortest" memory paths for that task. This might help.

There is a free beta test Windows OS tool that stops windows from constantly moving the threads around trying to get things "faster". This is a Windows design issue. It apparently doesn't do more than two cpu equivalents very well. Your 2990wx has the equivalent of 4 cpus in there.
It is here:
https://www.extremetech.com/computing/283114-new-utility-can-double-amd-threadripper-2990wx-performance

I would be VERY interested to see if the above helps. I didn't find out about this until after I had sold my 2990wx.

HTH,
Tom
I will stop procrastinating tomorrow.
\\// Live Long & Prosper (starting tomorrow ;)
ID: 1976602 · Report as offensive     Reply Quote
Profile Tom M
Volunteer tester

Send message
Joined: 28 Nov 02
Posts: 3380
Credit: 175,082,429
RAC: 813,127
United States
Message 1976643 - Posted: 23 Jan 2019, 20:18:42 UTC - in response to Message 1976413.  


- Auto snooze leading to down-clocking? (might not be applicable, and may depend on the motherboard)


If there is a "cool 'n quiet" in the bios it should be disabled. That is the AMD equivalent to Intel's speedstep parameter.

Tom
I will stop procrastinating tomorrow.
\\// Live Long & Prosper (starting tomorrow ;)
ID: 1976643 · Report as offensive     Reply Quote
Profile Keith Myers Special Project $250 donor
Volunteer tester
Avatar

Send message
Joined: 29 Apr 01
Posts: 9290
Credit: 829,202,446
RAC: 1,902,490
United States
Message 1976998 - Posted: 25 Jan 2019, 21:01:33 UTC - in response to Message 1976643.  

It depends on the brand of motherboard what the vendor calls the old "Cool 'n Quiet'. On ASUS motherboards it is called PSS.

This mobo has "pss support" bios feature. Should i disable it when overclocking, or not? What is pss support?

performance supported state, i guess. another name for CnQ


Turn it off or your motherboard will aggressively downclock at every advantage. You can also turn off C-states to keep the cpu in C0 state when loaded. Reduce the total cpu loading so the AMD power saving features don't downclock and keep more cores in boosted state.
Seti@Home classic workunits:20,676 CPU time:74,226 hours
ID: 1976998 · Report as offensive     Reply Quote
NovaExclusives

Send message
Joined: 26 Jan 19
Posts: 1
Credit: 0
RAC: 0
United States
Message 1977231 - Posted: 26 Jan 2019, 23:21:29 UTC

Hello Everyone!

I came to basically post about my experience with my 2990WX!

I currently have it at 3.6Ghz for my permanent desired settings at 1.0675vCore LLC 3 and can do 5300 reliably in Cinebench R15.

Going to run F@H tonight to get an idea of PC performance for you all here, and help Berkley with the demands of programming for not only 32 cores, but the incoming 64 core Black Edition Threadripper CPU's we are eagerly waiting for over at http://www.reddit.com/r/ripintel

Thanks for the help with confirming that my vCores and running temperatures (54c under most load, will get to 66c in Prime95 Max Heat on a 48h run) are all good for the next few years, let's hope board VRM's hold up!
ID: 1977231 · Report as offensive     Reply Quote
Profile Keith Myers Special Project $250 donor
Volunteer tester
Avatar

Send message
Joined: 29 Apr 01
Posts: 9290
Credit: 829,202,446
RAC: 1,902,490
United States
Message 1977236 - Posted: 26 Jan 2019, 23:35:18 UTC - in response to Message 1977231.  

May I ask which motherboard you are using for your 2990WX?
Seti@Home classic workunits:20,676 CPU time:74,226 hours
ID: 1977236 · Report as offensive     Reply Quote
jsm

Send message
Joined: 1 Oct 16
Posts: 76
Credit: 31,978,050
RAC: 83,218
Isle of Man
Message 1979068 - Posted: 7 Feb 2019, 7:50:38 UTC - in response to Message 1977236.  

I thought I would update now as I have been trying various suggestions to improve the throughput. For the record the MB is Asus Rog Extreme and the 16GB memory is at 3000 DDR4 which monitoring under full 64 thread load is not anywhere near 100% utilised.
I have installed Coreprio and tried all the options there (DLM, no DLM, disassociator, none etc) without any improvement. Then purchased ProjectLasso at very reasonable price running with coreprio. Again no obvious improvement - perhaps a little. I have just followd Tom's suggestion of reducing the cores available via BOINC at 50% cpus. So far not much change but the operation of coreprio forcing processing into first 16 cores which have direct memory access is clearly shown in the graph of Lasso. It lights up the first row in vivd green and the second hardly at all.
Perhaps I should make it clear that the purpose of this testing is to run Seti on the three Ryzen AMD m/cs with only SETI running other than WIN10 housekeeping.
I will keep you updated and am happy to try out any suggestions as so far I am sorely disappointed!
ID: 1979068 · Report as offensive     Reply Quote
Profile Tom M
Volunteer tester

Send message
Joined: 28 Nov 02
Posts: 3380
Credit: 175,082,429
RAC: 813,127
United States
Message 1979103 - Posted: 7 Feb 2019, 15:49:51 UTC - in response to Message 1979068.  
Last modified: 7 Feb 2019, 15:50:48 UTC

I thought I would update now as I have been trying various suggestions to improve the throughput. For the record the MB is Asus Rog Extreme and the 16GB memory is at 3000 DDR4 which monitoring under full 64 thread load is not anywhere near 100% utilised.
I have installed Coreprio and tried all the options there (DLM, no DLM, disassociator, none etc) without any improvement. Then purchased ProjectLasso at very reasonable price running with coreprio. Again no obvious improvement - perhaps a little. I have just followd Tom's suggestion of reducing the cores available via BOINC at 50% cpus. So far not much change but the operation of coreprio forcing processing into first 16 cores which have direct memory access is clearly shown in the graph of Lasso. It lights up the first row in vivd green and the second hardly at all.
Perhaps I should make it clear that the purpose of this testing is to run Seti on the three Ryzen AMD m/cs with only SETI running other than WIN10 housekeeping.
I will keep you updated and am happy to try out any suggestions as so far I am sorely disappointed!


After taking a look at some of the latest processing speed results I think that you should use BOINC Manager local configuration to take the number of threads down to .40/40%/26 threads. Based on reading some articles and my Linux experience I think 26 threads is the optimum for maximum total production for a 2990wx/2970wx.

I think the processing speed of the cpu tasks will take a significant jump (aka drop to 1 hour more or less).

What speed are you running the cpu at? Any overclocking? Have you turned off the "cool n' quiet" and/or PSS? as well as the C-states? The cool 'n quiet / PSS(sp?) are other names for what intel calls speedstep. Seti processing speeds up if speedstep is disabled.

HTH,
Tom
I will stop procrastinating tomorrow.
\\// Live Long & Prosper (starting tomorrow ;)
ID: 1979103 · Report as offensive     Reply Quote
Profile Keith Myers Special Project $250 donor
Volunteer tester
Avatar

Send message
Joined: 29 Apr 01
Posts: 9290
Credit: 829,202,446
RAC: 1,902,490
United States
Message 1979603 - Posted: 9 Feb 2019, 20:11:21 UTC - in response to Message 1979068.  

For the record the MB is Asus Rog Extreme and the 16GB memory is at 3000 DDR4

Where did you find 4GB 3000 DDR4 sticks? I can ever only find 2666 DDR4 in 4GB capacities.

Or did you only install 2 X 8GB sticks of memory. If that's the case, you are severely constraining the memory operating in 2 channel mode instead of the proper 4 channel mode.
Seti@Home classic workunits:20,676 CPU time:74,226 hours
ID: 1979603 · Report as offensive     Reply Quote
Kevray Project Donor

Send message
Joined: 29 Oct 99
Posts: 4
Credit: 19,708,305
RAC: 391
Australia
Message 1979677 - Posted: 10 Feb 2019, 7:17:46 UTC - in response to Message 1979103.  

I followed Toms advice & dropped the cores (24) in my 28970wx down to 40% & noticed an immediate improvement in speed. My Astropulse work unit time dropped from 24 hours + per unit down to about 6 hours.
I had improvements across other work units as well. I also use coreprio & im not sure if that helps as well. I've read the both Amd & Intel are working on a solution.
ID: 1979677 · Report as offensive     Reply Quote
jsm

Send message
Joined: 1 Oct 16
Posts: 76
Credit: 31,978,050
RAC: 83,218
Isle of Man
Message 1979694 - Posted: 10 Feb 2019, 10:23:42 UTC - in response to Message 1979603.  

The specialist builders installed it as 2 x 8gb when I asked for 16gb total. Are you saying that it is mandatory to have 4 channels of memory? I did not see this anywhere during reseach for my spec? If it is very significant should I lash out for two more 8gb sticks?
I have been running at 50% for three days and will reduce it to 40% today then knock off PSS on Wednesday. There is some improvement but to just about the same throughput as the 1800X m/cs which were less than a 1/3 of the cost.
jsm
ID: 1979694 · Report as offensive     Reply Quote
jsm

Send message
Joined: 1 Oct 16
Posts: 76
Credit: 31,978,050
RAC: 83,218
Isle of Man
Message 1979695 - Posted: 10 Feb 2019, 10:29:00 UTC - in response to Message 1979677.  

You have an average of 51,000 on Seti with a 24 core m/c while my 32 core unit is only managing 13,000. Something is BADLY wrong!
jsm
ID: 1979695 · Report as offensive     Reply Quote
Profile Tom M
Volunteer tester

Send message
Joined: 28 Nov 02
Posts: 3380
Credit: 175,082,429
RAC: 813,127
United States
Message 1979714 - Posted: 10 Feb 2019, 14:35:50 UTC - in response to Message 1979694.  

The specialist builders installed it as 2 x 8gb when I asked for 16gb total. Are you saying that it is mandatory to have 4 channels of memory? I did not see this anywhere during reseach for my spec? If it is very significant should I lash out for two more 8gb sticks?
I have been running at 50% for three days and will reduce it to 40% today then knock off PSS on Wednesday. There is some improvement but to just about the same throughput as the 1800X m/cs which were less than a 1/3 of the cost.
jsm


If you want to take advantage of the speedup of quad channel vs. dual channel memory you will need 4 sticks.

Tom
I will stop procrastinating tomorrow.
\\// Live Long & Prosper (starting tomorrow ;)
ID: 1979714 · Report as offensive     Reply Quote
Profile Tom M
Volunteer tester

Send message
Joined: 28 Nov 02
Posts: 3380
Credit: 175,082,429
RAC: 813,127
United States
Message 1979715 - Posted: 10 Feb 2019, 14:37:51 UTC - in response to Message 1979695.  
Last modified: 10 Feb 2019, 14:43:52 UTC

You have an average of 51,000 on Seti with a 24 core m/c while my 32 core unit is only managing 13,000. Something is BADLY wrong!
jsm


If you have just changed to parameters that show faster processing time it will take at least a week to see major changes in your RAC since it is an "average".

Check your bios. You are looking for the equivalent of Intel's speedstep that you want to disable. It is sometimes labeled "cool n' quiet" and sometimes labeled PSS. It depends on the brand of the MB. It could have another label. Might have to google the motherboard to find it.

Do this on both the Ryzen systems and the Threadripper. It makes a huge difference on cpu task processing speed.

Tom
I will stop procrastinating tomorrow.
\\// Live Long & Prosper (starting tomorrow ;)
ID: 1979715 · Report as offensive     Reply Quote
Profile Tom M
Volunteer tester

Send message
Joined: 28 Nov 02
Posts: 3380
Credit: 175,082,429
RAC: 813,127
United States
Message 1979716 - Posted: 10 Feb 2019, 14:54:53 UTC - in response to Message 1976998.  

It depends on the brand of motherboard what the vendor calls the old "Cool 'n Quiet'. On ASUS motherboards it is called PSS.

This mobo has "pss support" bios feature. Should i disable it when overclocking, or not? What is pss support?

performance supported state, i guess. another name for CnQ


Turn it off or your motherboard will aggressively downclock at every advantage. You can also turn off C-states to keep the cpu in C0 state when loaded. Reduce the total cpu loading so the AMD power saving features don't downclock and keep more cores in boosted state.


We have the experience that no BOINC Manager machine should run more than 90% of the cores. If you do, it slows production on the whole system.
I will stop procrastinating tomorrow.
\\// Live Long & Prosper (starting tomorrow ;)
ID: 1979716 · Report as offensive     Reply Quote
Profile Keith Myers Special Project $250 donor
Volunteer tester
Avatar

Send message
Joined: 29 Apr 01
Posts: 9290
Credit: 829,202,446
RAC: 1,902,490
United States
Message 1979726 - Posted: 10 Feb 2019, 16:04:42 UTC - in response to Message 1979694.  

The specialist builders installed it as 2 x 8gb when I asked for 16gb total. Are you saying that it is mandatory to have 4 channels of memory? I did not see this anywhere during reseach for my spec? If it is very significant should I lash out for two more 8gb sticks?
I have been running at 50% for three days and will reduce it to 40% today then knock off PSS on Wednesday. There is some improvement but to just about the same throughput as the 1800X m/cs which were less than a 1/3 of the cost.
jsm

Depends on the CPU architecture and the workload. The memory transfer bandwidth will be halved with only two sticks. No questions there and impossible to avoid. AMD architecture, especially Threadripper will be most impacted, Intel not so much depending on the workload.

Also the best way to speed up Threadripper is to run the memory interleaving in Channel mode to reduce memory latency. With only two sticks, you can't do that. So you will be stuck with 100ns latency instead of 60ns latency if you could run Channel mode interleaving.
Seti@Home classic workunits:20,676 CPU time:74,226 hours
ID: 1979726 · Report as offensive     Reply Quote
Profile Keith Myers Special Project $250 donor
Volunteer tester
Avatar

Send message
Joined: 29 Apr 01
Posts: 9290
Credit: 829,202,446
RAC: 1,902,490
United States
Message 1979727 - Posted: 10 Feb 2019, 16:10:44 UTC

The best memory to use on Ryzen/Threadripper is made with Samsung B-dies. Also the cpu speed is directly impacted by the speed of the memory. The faster you can run your memory, the faster the cpu will be. Samsung B-die memory runs very easily at 3466-3600 Mhz with low CL14 latencies. Lesser grade memory sticks made with Hynix or Micron memory dies usually only manages 2800-3200 Mhz at best with CL16 latencies.
Seti@Home classic workunits:20,676 CPU time:74,226 hours
ID: 1979727 · Report as offensive     Reply Quote
Profile Tom M
Volunteer tester

Send message
Joined: 28 Nov 02
Posts: 3380
Credit: 175,082,429
RAC: 813,127
United States
Message 1979773 - Posted: 10 Feb 2019, 21:46:57 UTC
Last modified: 10 Feb 2019, 21:47:43 UTC

I was going to assemble and bring up an Ryzen 7 2700 this week anyway as well as trying to get my 2400G box back up. So the current plan is to run both the 2400G which was already doing windows and the 2700 with windows.

Then I will have a better feel for trying to run a Ryzen 7 under windows with the stock seti applications.

I know that I never got my Ryzen 5 2400G cpu's tasks reliably down near an hour or so. It simply may not be possible with windows on the ryzen5/ryzen7.

I am looking forward to seeing Threadripper 2990wx CPU times dropping and the RAC climbing on that machine.

Tom
I will stop procrastinating tomorrow.
\\// Live Long & Prosper (starting tomorrow ;)
ID: 1979773 · Report as offensive     Reply Quote
1 · 2 · 3 · 4 . . . 29 · Next

Message boards : Number crunching : Ryzen and Threadripper


 
©2019 University of California
 
SETI@home and Astropulse are funded by grants from the National Science Foundation, NASA, and donations from SETI@home volunteers. AstroPulse is funded in part by the NSF through grant AST-0307956.