There's some interesting benchmarks posted for this over on Phoronix:

=========snip===========
Quite an intriguing mix of results. Threadripper wins outright on lower costs as is to be expected, but also does rather well for performance... See for yourself!

(As always, good comments in the forums comments for those articles...)

Happy cool crunchin',
Martin

I am not sure I saw any benchmarks that stress the cpu's the way Seti or other BOINC projects do.
So I would still be interested in running Windows 10 with/without the "fix" to see if it really can make a difference. My impression is, inspite the bench marking reported with an EPYC CPU, that there is something going on that explains the slowing performance. Unfortunately, once I came to the conclusion I could get more BOINC cpu processing speed out of a 2950wx there was no reason to keep my 2990wx.

I still suspect it is the asymmetrical memory system because the memory access bench marks show a difference in access speed. What I don't think we have is good instrumentation on is "memory path" congestion. We have some theories and we have some empirical evidence of over all slowdown. What we don't have is some way to directly measure the memory path congestion. I think. If I am wrong please let me know. Lots of this stuff fly's right over my head.

Tom

---

A proud member of the OFA (Old Farts Association).

correct. also thread ripper strange with memory even current stuff.

---

Tom I have a T/R 2970wx with 4 1070ti gpus,It has 48gb ram , water cooling & 1200 psu & appears to run very slow compared to my I 8700k (overclocked) with 2 1070ti gpus 16gb ram & no water cooling.
I have followed your thread & am aware of your & others problems with the memory issue.
Can you throw some light on what the problem is.

---

Tom I have a T/R 2970wx with 4 1070ti gpus,It has 48gb ram , water cooling & 1200 psu & appears to run very slow compared to my I 8700k (overclocked) with 2 1070ti gpus 16gb ram & no water cooling.
I have followed your thread & am aware of your & others problems with the memory issue.
Can you throw some light on what the problem is.

It depends on who you ask and what OS. My experience is with Linux/CUDA91,

Since you are running Windows my experience may not be applicable. But you are not the only person who is running a 2990wx under windows and it is running "very slow." My results were with the cpu running OC at about 3.6GHz because I had stability issues at speeds above that.

My testing is consistent with some other results that I have found. Basically, 26 cpu threads will minimize the time many cpu tasks take to run. This is with the 32c/64t 2990wx.

So 26/64 = 40%.so 40% of 48 = 19-20 threads.

I would try the following. Set your local BOINC Manager cpu configuration to 40% of the available cores and see what happens.

I am almost certain if you set it much higher than 40% your processing speed will slow down. If you set the amount of cpu used per gpu below 50% in your app_config.xml file or even don't have an app_config.xml file your gpus cpu core usage won't count towards the 40%/20 threads you are using.

I expect your average CPU task time to sink down to under 2 hours or lower. I haven't looked but I expect you are currently getting outlier cpu times of upto 6 hours per task. It will take time, maybe weeks for your average cpu time to sink down into the lower range. But you will "see" a lot more tasks when they hit 90%+ will have run times near or below 2 hours.

If you look on the leaderboard you will find Kevin Olley (sp?) who is running a 2990wx. He reports maximum production at 24 threads.

Under "stock" or even the anonymous distro for Windows, you should be able to speed up your gpu's some by using the command line file(s) for each gpu task.

If you need advice on setting the command line parameters for the Windows gpu tasks, just ask :)

Tom

---

A proud member of the OFA (Old Farts Association).

Thats great already within 15 mins I can see the difference, I have set the Cpu at 40% and I have 19 threads operating but the speed of processing is rising quickly.
I previously had a go at the app config to increase my Gpu crunching but it didnt take, dont know why?

---

Thats great already within 15 mins I can see the difference, I have set the Cpu at 40% and I have 19 threads operating but the speed of processing is rising quickly.
I previously had a go at the app config to increase my Gpu crunching but it didnt take, dont know why?

Exactly what do you mean by "increase my gpu crunching"? With the standard SOG type gpu task it might show some difference but not a lot.

Something like: " -sbs 1024 -period_iterations_num 10 -spike_fft_thresh 4096 -tune 1 64 1 4 -oclfft_tune_gr 256 -oclfft_tune_lr 16 -oclfft_tune_wg 256 -oclfft_tune_ls 512 -oclfft_tune_bn 64 -oclfft_tune_cw 64" in the mb*SOG.txt command line and restart the BoincManager might speed them up.

I used to get 7 minutes or so per SOG gpu task with a gtx 1060 3GB under Windows while using the above command line. With gtx 1070Ti you should do better.

Tom

---

A proud member of the OFA (Old Farts Association).