ended Fight Covid-19 & Win Prizes ! (Folding@Home) **Starts Midnight 08/04/20**

[tictoc]

Ok guys, anyone out there able to help me setup and optimize Threadripper 3970X? It looks like some people are getting it up to 1 million PPD...which seems like alot for a CPU??

 

I did a bit of googling and it seems there are some tweaks that need to happen to get the max out of this thing. I read some stuff about 'slot' settings for core/thread count but I am totally clueless.

 

Also, can anyone give me a really quick summary of how this scales with overclocking? For instance, how much system memory does it use and what the score increase? Overclocking CPU? Mem? I've been in the professional XOC world for some years now but never touched folding until yesterday...I know nothing.

 

Not many tweaks to do, but you will want to adjust the number of cores threads per CPU slot. You can do that using FAHControl. There are a few things that GROMACS (the program that is used behind the scenes for CPU folding) doesn't play nicely with. Primarily this is any thread count above 32 and also any thread count that is a prime > 7 or has a large prime factor. Also, it is best to set up multiple CPU slots, so that you will get the most work. What can happen if you have too high of a thread count is that you will get assigned work that does not use all of the available threads. I have found that the best max number of threads for each slot is 24. Any higher (up to 32) and you can still get work, but sometimes the WUs will be less than 32 threads.

 

Not sure how many GPUs you have in the system, but each NVIDIA GPU will need to be given at least 1 thread (I recommend 2 (one whole core)). Here is how I would set up a 32 core processor with two GPUs.

First CPU slot - 16

Second CPU slot - 16

Third CPU slot - 16

Fourth CPU slot - 10

That leaves a few cores to feed the GPUs and a core for general overhead.

 

To set the number of cores per slot go to Configure -> Slots

• Highlight your CPU slot and then click edit
  
• The default is -1, and that will just set it to fold on all threads -1.
  
• Set that to 16 then click OK.
  
• Now click on add
  
• Select CPU and set it to 16
  
• Do the same thing again until you have added all the slots you want and then click Save
  
• That should be it, and you should start to pick up work for each slot
  

Also for anyone else that is folding on their CPU, all of the new tasks, which are a_7 type work units, do use AVX instructions. This means you are going to see quite a bit more heat and power usage.

 

 

Nice !

 

Trying to see if I can get my CPU Folding stable (3950X) But is just does not like it. All other benchmarks pass including FAHBench lol, but whatever WU they are giving me on the CPU just blows up my OC.

 

The old version of FAHBench is compiled with an older type of CPU task that does not use AVX instructions.

 

Edited April 9, 2020 by tictoc

[ENTERPRISE]

 

Not many tweaks to do, but you will want to adjust the number of cores threads per CPU slot. You can do that using FAHControl. There are a few things that GROMACS (the program that is used behind the scenes for CPU folding) doesn't play nicely with. Primarily this is any thread count above 32 and also any thread count that is a prime > 7 or has a large prime factor. Also, it is best to set up multiple CPU slots, so that you will get the most work. What can happen if you have too high of a thread count is that you will get assigned work that does not use all of the available threads. I have found that the best max number of threads for each slot is 24. Any higher (up to 32) and you can still get work, but sometimes the WUs will be less than 32 threads.

 

Not sure how many GPUs you have in the system, but each NVIDIA GPU will need to be given at least 1 thread (I recommend 2 (one whole core)). Here is how I would set up a 32 core processor with two GPUs.

First CPU slot - 16

Second CPU slot - 16

Third CPU slot - 16

Fourth CPU slot - 10

That leaves a few cores to feed the GPUs and a core for general overhead.

 

To set the number of cores per slot go to Configure -> Slots

• Highlight your CPU slot and then click edit
  
• The default is -1, and that will just set it to fold on all threads -1.
  
• Set that to 16 then click OK.
  
• Now click on add
  
• Select CPU and set it to 16
  
• Do the same thing again until you have added all the slots you want and then click Save
  
• That should be it, and you should start to pick up work for each slot
  

Also for anyone else that is folding on their CPU, all of the new tasks, which are a_7 type work units, do use AVX instructions. This means you are going to see quite a bit more heat and power usage.

 

 

 

The old version of FAHBench is compiled with an older type of CPU task that does not use AVX instructions.

 

Brill advice for @mllrkllr88 bud. I actually remember back in the day now that you had to leave a Core to each GPU so it can manage the GPU things correctly when folding. If I get my CPU stable I will setup my folding client correctly this time for the CPU. Out of interest, if you do not assign a core to the GPU when CPU and GPU folding, I would imagine that just causes lower performance, not a system crash ?

[Simmons]

 

Brill advice for @mllrkllr88 bud. I actually remember back in the day now that you had to leave a Core to each GPU so it can manage the GPU things correctly when folding. If I get my CPU stable I will setup my folding client correctly this time for the CPU. Out of interest, if you do not assign a core to the GPU when CPU and GPU folding, I would imagine that just causes lower performance, not a system crash ?

 

If you just let the system max out, and you don't spare a core for the GPU, the biggest thing that will happen is that the CPU will be splitting the load between the CPU WU and the GPU WU. Basically, having the extra free core allows you some headroom to ensure maximum PPD on the GPU.

 

So yeah, no crash.

[ENTERPRISE]

 

If you just let the system max out, and you don't spare a core for the GPU, the biggest thing that will happen is that the CPU will be splitting the load between the CPU WU and the GPU WU. Basically, having the extra free core allows you some headroom to ensure maximum PPD on the GPU.

 

So yeah, no crash.

 

Yeah I thought as much. Going to test my OC with some hefty AVX workloads/benches to see if I can catch an instability. Ran two runs of Memtest, no memory instabilities were shown there as both tests were error free + other memory related benchmarks pass without issue. Will also try dropping to 4.3Ghz as well to see if that makes any difference.

Edited April 10, 2020 by ENTERPRISE

[bigblueshock]

Hi Guys!

 

I'm new here, but will be assisting the ExtremeHW team for folding at home for the Covid 19 Fight I've already started folding last night.

 

I am using all cores on my AMD Ryzen 3900x as well as my GeForce 1080 Ti with a voltage limit of .875 (when folding, equivalent to about 65% card power usage)

 

I do have a recommendation for those on Ryzen/Threadripper 3rd gen. Use 1usmus DRAM Calculator. Not only can it help assist you with Memory Timings/Voltages, but it also has a MEMBench /memtest built in.

 

Under the MEMbench tab I recommend Memtest mode with a Task Scope of at least 200%. Depending on your memory speed/timings, I would let it run at least 30-40 minutes. If there are any issues with your settings/voltages, it will stop on 1st error by default.

[Simmons]

Hi Guys!

 

I'm new here, but will be assisting the ExtremeHW team for folding at home for the Covid 19 Fight I've already started folding last night.

 

I am using all cores on my AMD Ryzen 3900x as well as my GeForce 1080 Ti with a voltage limit of .875 (when folding, equivalent to about 65% card power usage)

 

I do have a recommendation for those on Ryzen/Threadripper 3rd gen. Use 1usmus DRAM Calculator. Not only can it help assist you with Memory Timings/Voltages, but it also has a MEMBench /memtest built in.

 

Under the MEMbench tab I recommend Memtest mode with a Task Scope of at least 200%. Depending on your memory speed/timings, I would let it run at least 30-40 minutes. If there are any issues with your settings/voltages, it will stop on 1st error by default.

 

Hey and welcome!

 

Would you show us a screenshot of your folding client configuration? Since you have that many cores, you may need to optimize the CPU slots a bit.

[bigblueshock]

 

Hey and welcome!

 

Would you show us a screenshot of your folding client configuration? Since you have that many cores, you may need to optimize the CPU slots a bit.

 

Much appreciated. This is my first time using F@H in over 10 years so I'm sure my config can be optimized. Currently I've left everything at default with the exception of Power Consumption set to Full.

 

I tried attaching three JPG's but it's saying blocked by file scanner, so I'm sharing via three image links.

 

https://ibb.co/xMB8S6V

https://ibb.co/SK3m629

https://ibb.co/dc6m7VM

 

Thank you!

[Simmons]

Okay so the big thing I will suggest is manually splitting your CPU into 2 slots: 1 slot with 16 threads, and 1 slot with 6 threads. This would reserve a couple threads for your GPU.

 

Also, you may want to add the flag "next-unit-percentage" with a value of "100" for all of your slots. It seems to make grabbing WUs more reliable these days, espeially with the WU drought we've been experiencing.

[ENTERPRISE]

 

Much appreciated. This is my first time using F@H in over 10 years so I'm sure my config can be optimized. Currently I've left everything at default with the exception of Power Consumption set to Full.

 

I tried attaching three JPG's but it's saying blocked by file scanner, so I'm sharing via three image links.

 

https://ibb.co/xMB8S6V

https://ibb.co/SK3m629

https://ibb.co/dc6m7VM

 

Thank you!

 

Apologies for that. We are working on our file AV which caused this, I have turned it off for now. However for images, please use our image up-loader as per the red camera icon to the top right of the editor Or you can continue linking. Welcome to EHW and the Fold !

[bigblueshock]

Okay so the big thing I will suggest is manually splitting your CPU into 2 slots: 1 slot with 16 threads, and 1 slot with 6 threads. This would reserve a couple threads for your GPU.

 

Also, you may want to add the flag "next-unit-percentage" with a value of "100" for all of your slots. It seems to make grabbing WUs more reliable these days, espeially with the WU drought we've been experiencing.

 

Thank you!! Does this look right? https://ibb.co/D1TQrZS

 

I now have three total slots. One with 16 threads, one with 6 threads, and one for my GPU

[Simmons]

Looks good! You should be good to go once you get a few WUs :thumbs_thumbup:

[bigblueshock]

Looks good! You should be good to go once you get a few WUs :thumbs_thumbup:

 

?

 

That's actually another question I had. Probably a simple answer. Is there a correlation between points and WU's? I noticed I have more WU's (22) than people with more points than me.

[Supercrumpet]

 

?

 

That's actually another question I had. Probably a simple answer. Is there a correlation between points and WU's? I noticed I have more WU's (22) than people with more points than me.

 

This may not be completely accurate, but historically I've seen more powerful cards tend to get bigger (higher point award) WUs.

[Alex]

Been getting a fair amount of WUs on my CPU, not many on my GPU

[Simmons]

Been getting a fair amount of WUs on my CPU, not many on my GPU

 

Same, it seems like there are less available today for whatever reason. You'd think more people would be folding during the week, rather than on a Friday lol

[bigblueshock]

 

Same, it seems like there are less available today for whatever reason. You'd think more people would be folding during the week, rather than on a Friday lol

 

100% True, other than the fact everyone's either working from home, or not working right now. lol

[ENTERPRISE]

So it looks like I am having a much better success at 4.3Ghz on the 3950X when folding. Will get it stable folding at the lowers Vcore I can and then run some other benches to make sure it is rock solid, then I can safely commit the CPU to the cause to, woop !

[bigblueshock]

So it looks like I am having a much better success at 4.3Ghz on the 3950X when folding. Will get it stable folding at the lowers Vcore I can and then run some other benches to make sure it is rock solid, then I can safely commit the CPU to the cause to, woop !

 

If you're thinking of lowering vcore, or even figuring out how it works... I would take a look at this review: https://www.gamersnexus.net/guides/3...ffset-override

 

You can actually lower vCore all you want, even go as low as 1v probably. Nothing will crash, however you will actually sacrifice major performance. These new chips work very differently than 1000 and 2000 series Ryzen chips. You will literally have to run something line Cinebench for each voltage increment until you know you haven't lost performance.

 

I've had my 3900x since last August and read/watched tons of videos regarding how it works, including dos and donts. The general consensus is don't even bother trying to manually OC it for 24/7 use. The performance gain is minimal if any, and you lose the single core / multi-core higher boost.

 

I suppose if If you're onlyyyy doing Folding at Home and you're on a custom loop and always maximizing the use of all threads, then maybe you could play around with manual/static OC But it's going to be a pain...

 

Edit: I should have added, If you keep everything stock (minus the RAM of of course, that always has to be manually adjusted), the CPU will adjust frequency based on CPU temperature. AMD's algorithm works well unless you're running liquid nitrogen.

 

For instance, I've gotten higher Cinebench scores in the Winter time, and slightly lower scores in the Summer time (all due to ambient temperatures).

 

Good luck!

Edited April 10, 2020 by bigblueshock

[tictoc]

Manual 4.3 OC here on my 3970X 3960X, but my CPU only gets a short break once every few weeks when I have to reboot for some updates.

For my situation where the CPU is always at a minimum of 90% load, the overshoot/undershoot voltage swings with auto-boost and PBO were never really stable.

 

Edited April 10, 2020 by tictoc

[bigblueshock]

Manual 4.3 OC here on my 3970X, but my CPU only gets a short break once every few weeks when I have to reboot for some updates.

For my situation where the CPU is always at a minimum of 90% load, the overshoot/undershoot voltage swings with auto-boost and PBO were never really stable.

 

That makes sense the way you're using the chip then. Yeah, as far as I'm concerned, PBO is still kind of broken anyhow. I actually got lower scores with PBO enabled, lol.

[ENTERPRISE]

 

If you're thinking of lowering vcore, or even figuring out how it works... I would take a look at this review: https://www.gamersnexus.net/guides/3...ffset-override

 

You can actually lower vCore all you want, even go as low as 1v probably. Nothing will crash, however you will actually sacrifice major performance. These new chips work very differently than 1000 and 2000 series Ryzen chips. You will literally have to run something line Cinebench for each voltage increment until you know you haven't lost performance.

 

I've had my 3900x since last August and read/watched tons of videos regarding how it works, including dos and donts. The general consensus is don't even bother trying to manually OC it for 24/7 use. The performance gain is minimal if any, and you lose the single core / multi-core higher boost.

 

I suppose if If you're onlyyyy doing Folding at Home and you're on a custom loop and always maximizing the use of all threads, then maybe you could play around with manual/static OC But it's going to be a pain...

 

Edit: I should have added, If you keep everything stock (minus the RAM of of course, that always has to be manually adjusted), the CPU will adjust frequency based on CPU temperature. AMD's algorithm works well unless you're running liquid nitrogen.

 

For instance, I've gotten higher Cinebench scores in the Winter time, and slightly lower scores in the Summer time (all due to ambient temperatures).

 

Good luck!

 

Thanks for the info. It was an interesting read, including the GN article. I am not aiming for a super low vcore, just a sensible minimum that will keep my all core OC stable. Much of the work or software I use is multi-threaded, so it makes more sense for me to have a manual maximum all core OC over an auto OC over PBO which will only give a favourable boost on certain cores. PBO is fine for sure, just depends on your use case. I may do some testing with PBO again after I re-find my stable point with Folding with respect to vcore.

 

Will be interesting if they have improved PBO at all.

[tictoc]

 

?

 

That's actually another question I had. Probably a simple answer. Is there a correlation between points and WU's? I noticed I have more WU's (22) than people with more points than me.

 

If you are folding on your CPU, the ratio of WU-points will be lower than GPU folding. The reason for this is twofold:

• Current CPU WUs are fairly small with a low TPF (time per frame) so they will complete quickly
  
• Much of protein folding can be highly parallelized, so with an optimized GPU application, GPUs are able to do more work in a shorter amount of time than CPUs
  

Additionally, faster GPUs and CPUs will see more points per WU. This is because of QRB (quick return bonus) which awards more points the faster that the work is turned around from server-client-server. With the newer CPU WUs, it is not nearly as big of a divide as it was a few years ago before the newer a7 tasks were released.

 

[FreeJet]

Shared on Twitter:

 

[bigblueshock]

 

If you are folding on your CPU, the ratio of WU-points will be lower than GPU folding. The reason for this is twofold:

• Current CPU WUs are fairly small with a low TPF (time per frame) so they will complete quickly
  
• Much of protein folding can be highly parallelized, so with an optimized GPU application, GPUs are able to do more work in a shorter amount of time than CPUs
  

Additionally, faster GPUs and CPUs will see more points per WU. This is because of QRB (quick return bonus) which awards more points the faster that the work is turned around from server-client-server. With the newer CPU WUs, it is not nearly as big of a divide as it was a few years ago before the newer a7 tasks were released.

 

This makes perfect sense. Thanks!

[ENTERPRISE]

Nice.

 

At the time of writing this post, 351 Work Units have been completed !