Folding Community Board

[leadeater]

52 minutes ago, RollinLower said:

so i should be splitting it up into 3 then
i wonder how that is handled within the chip, what CCXs are handling what workload etc. i guess it would be beneficial to have as much of the same chiplets working on a single WU as possible.

You should make slots based on the CPU architecture, memory NUMA domains and also sub-NUMA domains for caches/core. I didn't really optimize at all for last event, I just made multiple 32 thread slots and filled up all 192 threads of the 4x 8890v4 (24/48). I should have been creating 24 thread slots based on these CPUs.

 

Core Zen you should create the slots no bigger than a CCX/CCD I suspect, I have not got around to actually testing this but it seems the most logical/optimal.

 

config.xml

<config>
  <!-- Client Control -->
  <fold-anon v='true'/>

  <!-- Folding Slot Configuration -->
  <gpu v='false'/>

  <!-- Slot Control -->
  <power v='full'/>

  <!-- User Information -->
  <passkey v='*****'/>
  <team v='223518'/>
  <user v='******'/>

  <!-- Folding Slots -->
  <slot id='0' type='CPU'>
    <cpus v='32'/>
  </slot>
  <slot id='1' type='CPU'>
    <cpus v='32'/>
  </slot>
  <slot id='2' type='CPU'>
    <cpus v='32'/>
  </slot>
  <slot id='3' type='CPU'>
    <cpus v='32'/>
  </slot>
  <slot id='4' type='CPU'>
    <cpus v='32'/>
  </slot>
  <slot id='5' type='CPU'>
    <cpus v='32'/>
  </slot>
  <slot id='6' type='CPU'>
    <cpus v='32'/>
  </slot>

  <allow v='****' />
  <web-allow v='****' />
</config>

 

Having 2 systems with the same CPU really helps figure out what is best, can try 2 slightly different configurations and see which comes out as the best PPD over a few days.

[RollinLower]

9 hours ago, leadeater said:

 

Core Zen you should create the slots no bigger than a CCX/CCD I suspect, I have not got around to actually testing this but it seems the most logical/optimal.

So according to what I could find online, this chip should have 4 CCDs made up of 8 core complexes with 6 cores each. Which would mean I should make 4 slots of 12 cores(24threads) each to optimally spread the load. Does that make sense?

[leadeater]

4 hours ago, RollinLower said:

So according to what I could find online, this chip should have 4 CCDs made up of 8 core complexes with 6 cores each. Which would mean I should make 4 slots of 12 cores(24threads) each to optimally spread the load. Does that make sense?

Since it's Zen 2 I think the most optimal would be Slot size of 3 threads, but that's a bit tricky of a recommendation and it also depends how good the thread scheduling is.

 

What you generally want to avoid is inter-core traffic between CCX's and each CCX on that CPU is only 3 cores and 6 threads so the maximum I would do is 6 however not all 6 may end up on the same CCX but they should. So I would test Slot size of 6 and 12 totaling to 96.

 

FAH only cares about thread count and Linux handles the rest, basically you either target 2 CPU Slots per CCD or 1 CPU Slot per CCD and configure the threads per Slot accordingly. Technically you could break it down to CCX but you'd be creating a lot of CPU slots and I have no idea how effect on performance that would have

• 32x 3 Thread Slots (should be the best, maybe)
• 16x 6 Thread Slots
• 8x 12 Thread Slots
• 4x 24 Thread Slots (should be the worst, maybe)

Really hard to say without knowing more about the actual workloads and we don't really get to know that much about them and there are different ones that get run anyway.

 

This is also why Zen 3 is so much better, 8 core CCX is really much better than 4 core CCX since communication between CCXs on the same CCD still have to travel back to the IOD then back to the same CCD but to the opposite CCX.

[GOTSpectrum]

17 hours ago, BondiBlue said:

Oops! I've been very busy lately, and I thought my office PC was still chugging along with F@h on the two GPUs it has. Looks like it went down a few days ago for some reason. It's back up and running now, but I lost out on a good amount of points. The GTX 1080 + RTX 2070 Super combo in that machine does really well for F@h. 

 

I had the same thing, but was home so noticed it 

[RollinLower]

5 hours ago, leadeater said:

Since it's Zen 2 I think the most optimal would be Slot size of 3 threads, but that's a bit tricky of a recommendation and it also depends how good the thread scheduling is.

 

What you generally want to avoid is inter-core traffic between CCX's and each CCX on that CPU is only 3 cores and 6 threads so the maximum I would do is 6 however not all 6 may end up on the same CCX but they should. So I would test Slot size of 6 and 12 totaling to 96.

 

FAH only cares about thread count and Linux handles the rest, basically you either target 2 CPU Slots per CCD or 1 CPU Slot per CCD and configure the threads per Slot accordingly. Technically you could break it down to CCX but you'd be creating a lot of CPU slots and I have no idea how effect on performance that would have

• 32x 3 Thread Slots (should be the best, maybe)
• 16x 6 Thread Slots
• 8x 12 Thread Slots
• 4x 24 Thread Slots (should be the worst, maybe)

Really hard to say without knowing more about the actual workloads and we don't really get to know that much about them and there are different ones that get run anyway.

 

This is also why Zen 3 is so much better, 8 core CCX is really much better than 4 core CCX since communication between CCXs on the same CCD still have to travel back to the IOD then back to the same CCD but to the opposite CCX.

i'm now running the worst setup in your post. i'll log the following WU's and then i'll move to a more optimal setup step by step. lets see what the difference will be! should be pretty interesting.

[justpoet]

9 hours ago, RollinLower said:

i'm now running the worst setup in your post. i'll log the following WU's and then i'll move to a more optimal setup step by step. lets see what the difference will be! should be pretty interesting.

Some anecdotal AMD CPU folding info.  I've found with my desktop parts (2400g, 3400g, 5700g), they all were best™ with 1 thread per physical core when setting up F@H.  Note though, that these are all either monolythic or in essence single CCD type designs.  If I go from 8 threads to 16 threads on my 5700g that I currently have, for example, I will gain 10k'ish PPD on some WUs and lose 15Kish PPD on some other WUs, while using more power and running hotter almost all the time.  I suspect that the hit would be less (maybe end up better) if I made it less secure to side channel attacks in EFI and the OS, but since folding is a side thing, not a main thing, despite very low to no CPU usage otherwise most of the time, I'm going to leave it as secure as possible.

[leadeater]

6 hours ago, justpoet said:

Some anecdotal AMD CPU folding info.  I've found with my desktop parts (2400g, 3400g, 5700g), they all were best™ with 1 thread per physical core when setting up F@H.  Note though, that these are all either monolythic or in essence single CCD type designs.  If I go from 8 threads to 16 threads on my 5700g that I currently have, for example, I will gain 10k'ish PPD on some WUs and lose 15Kish PPD on some other WUs, while using more power and running hotter almost all the time.  I suspect that the hit would be less (maybe end up better) if I made it less secure to side channel attacks in EFI and the OS, but since folding is a side thing, not a main thing, despite very low to no CPU usage otherwise most of the time, I'm going to leave it as secure as possible.

The APUs also have less L3 cache too, 1/4 for Ryzen 3000 compared to the mainstream desktop parts and 1/2 typically for Ryzen 5000.

[Gorgon]

22 hours ago, RollinLower said:

i'm now running the worst setup in your post. i'll log the following WU's and then i'll move to a more optimal setup step by step. lets see what the difference will be! should be pretty interesting.

You might find this thread this thread and this thread over at FoldingForm of interest.

[jctappel67]

Made it to team rank 45, and now due to financial concerns, I'll have to stop until Folding Month 2023. I'll be around. Watching. Waiting. Commiserating.

 

[RollinLower]

6 hours ago, jctappel67 said:

Watching. Waiting. Commiserating

that sounds like you're planning some kind of murder for next event.

[GOTSpectrum]

Well I'm on holiday this week and apparently the internet las been down at home since I left so no folding being done atm 

[Gorgon]

As Ubuntu 18.04 LTS is end-of-life this year I've updated my Linux Folding Rig HOWTO for Ubuntu 22.04:

 

[CT854]

I started noticing some artifacting with my rig a while back so I quit folding on it for the longest time. I've also been extremely busy with work and other life commitments so I haven't gotten another rig up and running for it.

 

I shudder to think just how many spots I fell on the leaderboard, but one day I'll claw my way back up......

[GOTSpectrum]

5 minutes ago, CT854 said:

I started noticing some artifacting with my rig a while back so I quit folding on it for the longest time. I've also been extremely busy with work and other life commitments so I haven't gotten another rig up and running for it.

 

I shudder to think just how many spots I fell on the leaderboard, but one day I'll claw my way back up......

rank 103 for you right now 

[RollinLower]

On 6/17/2023 at 7:22 PM, RollinLower said:

i'm now running the worst setup in your post. i'll log the following WU's and then i'll move to a more optimal setup step by step. lets see what the difference will be! should be pretty interesting.

and just like that, i forgot i was logging WU's.
i should have literal mountains of data on the worst setup now though

[RollinLower]

11 hours ago, RollinLower said:

and just like that, i forgot i was logging WU's.
i should have literal mountains of data on the worst setup now though

i also found that this 200Watt EPYC 7552 chip make my loop substantially hotter than the previous gen 180W 7551 i had. 
like, 15 degrees hotter on the cpu package? that seems excessive.

re-mounted the CPU block just now, but no dice. seems like these are just very toasty chips?

[jctappel67]

On 6/20/2023 at 3:14 PM, RollinLower said:

that sounds like you're planning some kind of murder for next event.

It was meant to be from "All the Small Things" by Blink-182, but I like your version better.

[Gorgon]

9 hours ago, micha_vulpes said:

...

Is anyone else seeing odd behavior ? I am not sure if this is a driver bug on version 535.98 DCH, some update to the way F@H is marking / running the workloads, or something else. You can see its just really tailed off in the last few days. F@H core 22 gets its own dedicated hardware thread, its running as high priority, etc etc.

 

Never mind, Fixed it. Migrated driver back to the Production branch (474.44 WHQL May 11 2023) appears to fix idle clock locking without needing a 3d render open.

 

Package power is a lot more stable, locked at about 138 watts only varying +/- 2 watts, vs the regular branch driver that would swing 25watts pretty consistently.

I'm not certain you'd have to revert back that far. I observed issues with 531 on Windows not reporting the actual Graphics Clock on a 3080 a month or so ago and reverted back to 528.49 and have had no issues since. My 4000-series (2 x 4070ti; 4070 & 4060ti) under Linux are all running 525.126 with no issues as were my 3000-series under Linux (2 x 3080; 3070ti & 3070) but those have been shut-down for the summer to reduce the extra load on the Air Conditioning during the hot weather.

I suspect the 52x branch is more stable than the 53x branch these days.

[Gorgon]

@GOTSpectrum did you finally get some time to tune your 4070ti and, if so, how is your yield?

[GOTSpectrum]

2 minutes ago, Gorgon said:

@GOTSpectrum did you finally get some time to tune your 4070ti and, if so, how is your yield?

Honestly, I just left it as is for the most part. Power limited to 80% and left it there, I may play with it at some point but I also use it for gaming so don't want to affect it too much 

[dogwitch]

turn off folding right. unless a driver update. for past 2 months. due to power bill. also power company gouging prices before new rate kick in next month

[BondiBlue]

Unfortunately I'm going to have to slow down on my folding activities for a bit. The power bill for the past month was getting dangerously close to $400. Only a small part of that was from my folding computers. Most of it was caused by my air conditioner. It's a central AC unit, but there's just the one unit for two floors. It's not quite large enough for both floors, and when it gets very hot outside (like it has been lately) that thing runs nonstop. Adding to the heat load by folding just isn't something I can do right now, so my numbers are going to be rather low for a while. My wallet will thank me for keeping the bill lower lol. 

[Mxyzptlk]

I have backed off as well… It’s expected with the summer heat.  

[GOTSpectrum]

It's proving to be mild here so I'm going full bore 

[GOTSpectrum]

Sneak peak, this design is not final but close