Folding Community Board

[Gorgon]

4 hours ago, CWallenberg said:

RTX 5080 returns "tbd" when Folding.LAR.systems webclient is looking up the GPU against it's AP I. Sample wu submit seems to also return an error

 

  Hide contents

 

 

 

@LAR_Systems

[Fallguye]

Wow just looking at the first few RTX 50xx series entries on LAR, they are surprisingly underwhelming. Though maybe I shouldn't be surprised given the poor performance in everything else...

 

The 5070 Ti is currently barely faster than a 4070 Super, and the 5090 is first but PPD/W is not really better than the 4090.

[gravitywave]

5 hours ago, Fallguye said:

Wow just looking at the first few RTX 50xx series entries on LAR, they are surprisingly underwhelming. Though maybe I shouldn't be surprised given the poor performance in everything else...

 

The 5070 Ti is currently barely faster than a 4070 Super, and the 5090 is first but PPD/W is not really better than the 4090.

I was surprised to see the numbers there on the 5070 Ti. Guess it's good the one I bought was defective and I had to return it. I really expected it to perform more around a 4080S.

[Gorgon]

11 hours ago, Fallguye said:

Wow just looking at the first few RTX 50xx series entries on LAR, they are surprisingly underwhelming. Though maybe I shouldn't be surprised given the poor performance in everything else...

 

The 5070 Ti is currently barely faster than a 4070 Super, and the 5090 is first but PPD/W is not really better than the 4090.

You have to remember that the 5000-series are manufactured on the same process node (TSMC 5nm) as the 4000-series. Nvidia has relegated their consumer GPUs to the older process nodes and is using their capacity on newer nodes for their Professional GPUs. So you should expect only slight improvements in F@H yields when comparing cards with similar number of CUDA cores.

 

The best strategy for folks looking to upgrade is likely to look for 4000-series cards on the used market from people upgrading to the 5000-series.

[muziqaz]

PSA: AMD decided to separate opencl-icd-loader package from ROCm 6.3.x, and it is no longer available from that repo. AMD hopes that distro repos would have that package. My first try with Mint Linux ended in failure. This is only applicable to 6.3 and later, but it is possible they might wipe the damn thing from older repos.

Why am I mentioning this here: fahclient relies on that package to install libOpenCL.so, which shows fahclient v8 that AMD opencl platform is available.

 

AMD opencl in Linux ONLY

[Fallguye]

1 hour ago, Gorgon said:

You have to remember that the 5000-series are manufactured on the same process node (TSMC 5nm) as the 4000-series. Nvidia has relegated their consumer GPUs to the older process nodes and is using their capacity on newer nodes for their Professional GPUs. So you should expect only slight improvements in F@H yields when comparing cards with similar number of CUDA cores.

 

The best strategy for folks looking to upgrade is likely to look for 4000-series cards on the used market from people upgrading to the 5000-series.

Oh I appreciate there are limits to what Nvidia can do without a die shrink, but in the past they have worked to improve architecture between generations. But the CUDA cores for the 50 and 40 series seem to be very similar in performance.

 

I guess maybe all their efforts were focused on "AI" enhancements that don't help folding or gaming. Or maybe there are some CUDA efficiency gains but they require code to be reworked to use them.

[Fallguye]

1 hour ago, muziqaz said:

PSA: AMD decided to separate opencl-icd-loader package from ROCm 6.3.x, and it is no longer available from that repo. AMD hopes that distro repos would have that package. My first try with Mint Linux ended in failure. This is only applicable to 6.3 and later, but it is possible they might wipe the damn thing from older repos.

Why am I mentioning this here: fahclient relies on that package to install libOpenCL.so, which shows fahclient v8 that AMD opencl platform is available.

 

AMD opencl in Linux ONLY

Thanks for the heads up, my adventures in Mint are still CPU only so not an issue for me yet.

[muziqaz]

32 minutes ago, Fallguye said:

Thanks for the heads up, my adventures in Mint are still CPU only so not an issue for me yet.

So for now ROCm 6.2.4 is a certainty as per this guide: https://foldingforum.org/viewtopic.php?t=42491
 

Unfortunately 6.2.x does not support RDNA4, as far as I know

[LAR_Systems]

On 3/10/2025 at 2:31 PM, muziqaz said:

That GPU entry in gpus.json is normal as other entries. Not sure why lar is picking it up like that

Because the gpus API at one point had TBD in it, like others that have Reserved ID etc. that I assume will change to a real device name.

I have added code to allow update of such entries in my DB after they are initially set, this should resolve the issue.

[muziqaz]

42 minutes ago, LAR_Systems said:

Because the gpus API at one point had TBD in it, like others that have Reserved ID etc. that I assume will change to a real device name.

I have added code to allow update of such entries in my DB after they are initially set, this should resolve the issue.

I'm not aware of any reservations. ID is either on the list or it is not. We do have some nVidia GPUs which are species 0, which means unsupported, but that's about it

[LAR_Systems]

21 hours ago, muziqaz said:

I'm not aware of any reservations. ID is either on the list or it is not. We do have some nVidia GPUs which are species 0, which means unsupported, but that's about it

 

The device IDs once added to the list might not change, but the descriptions appear to have placeholders sometimes that then become the official / final string format.

[muziqaz]

1 hour ago, LAR_Systems said:

 

The device IDs once added to the list might not change, but the descriptions appear to have placeholders sometimes that then become the official / final string format.

Oh, that's on nVidia. That's why I never noticed this. I can talk to the person who does nVidia (unless someone with access to nVidia hardware which is unreleased, add these IDs for testing purposes...). So what exactly your extension does not like in gpus.txt? Is it the change in the name (description)? Or is it changing from specific name to the final official name? I know for a fact that on AMD side I changed the names many many times. But they were always something to do with the GPU it represents.

[muziqaz]

So, basically let me know how you would prefer us to fill the ID's which have not been assigned a GPU to avoid messing with your system?

[LAR_Systems]

On 3/13/2025 at 7:49 PM, muziqaz said:

So, basically let me know how you would prefer us to fill the ID's which have not been assigned a GPU to avoid messing with your system?

Should not be a problem now.

 

If the F@H / Vendor changes the description due to reclassification or updating placeholders my side will pick up the update about an hour later after a change to the code I made on my side.

[muziqaz]

14 minutes ago, LAR_Systems said:

Should not be a problem now.

 

If the F@H / Vendor changes the description due to reclassification or updating placeholders my side will pick up the update about an hour later after a change to the code I made on my side.

Awesome stuff. Let me know if we need to change our habits on our end

 

[muziqaz]

Here is a guide to run AMD rx9070 on Linux Mint (and other distros) with rocm 6.3.4:
https://foldingforum.org/viewtopic.php?p=368095#p368095

Week of troubleshooting and working theories out brought the results

[NetoriusNick]

On 2/28/2025 at 7:35 PM, Shlouski said:

I've built myself a little server room, its well ventilated with postive pressure and equipped with air filters.

 

I have setup my FAH machine and I'm currently monitoring the temperatures, so far no missing ROP's or burnt power cables.

 

 

SIGH  If only if only 

[NetoriusNick]

On 3/6/2025 at 7:32 PM, muziqaz said:

I hope you know that nVidia paper launched this generation to the point where there are no cards to buy at all. 5090 was the exception, while still was a paper launch

One of the reason I'm not buying cards anymore, other is a new family () and no money

[Shlouski]

2 hours ago, NetoriusNick said:

new family

Congratulations, you might not have a lot of money but you are rich.

[Kazi]

Is there anyone who has experience with messing with GPUs.txt? 

 

fahclient --lspci showed me  
0x1414:0x008e:0:0:0:Microsoft Corporation:NVIDIA GeForce RTX 4080 SUPER

 

It’s partitioned GPU and it seems all synthetic benchmarks are able to utilize it, but only the Folding at home is not able to recognize it. 
My guess is it’s because of missing an entry in GPUs.txt.

 

Does anyone know how to trick the FAH client into recognize GPU which is not defined in GPUs.txt?

 

I don’t want to run it long term, I just want to know if it’s possible and how it will affect PPD when folding both on host and guest.  

 

EDIT: I also wanted to mention that FAH client sees the CUDA and OpenCL device 

[GOTSpectrum]

1 hour ago, Kazi said:

Is there anyone who has experience with messing with GPUs.txt? 

 

fahclient --lspci showed me  
0x1414:0x008e:0:0:0:Microsoft Corporation:NVIDIA GeForce RTX 4080 SUPER

 

It’s partitioned GPU and it seems all synthetic benchmarks are able to utilize it, but only the Folding at home is not able to recognize it. 
My guess is it’s because of missing an entry in GPUs.txt.

 

Does anyone know how to trick the FAH client into recognize GPU which is not defined in GPUs.txt?

 

I don’t want to run it long term, I just want to know if it’s possible and how it will affect PPD when folding both on host and guest.  

 

EDIT: I also wanted to mention that FAH client sees the CUDA and OpenCL device 

I'm fairly sure this is not a f@h issue, I've had similar issues before, even restarting the client, rebooting the PC and going as far as installing the client didn't fix it. 

It was always some kind of hardware/firmware/driver issue... 

Even as much as pulling the GPU, cleaning it and plugging it back in resolved it. Have you looked in the GPUs.txt file yet? 

[muziqaz]

On 3/18/2025 at 8:06 PM, Kazi said:

Is there anyone who has experience with messing with GPUs.txt? 

 

fahclient --lspci showed me  
0x1414:0x008e:0:0:0:Microsoft Corporation:NVIDIA GeForce RTX 4080 SUPER

 

It’s partitioned GPU and it seems all synthetic benchmarks are able to utilize it, but only the Folding at home is not able to recognize it. 
My guess is it’s because of missing an entry in GPUs.txt.

 

Does anyone know how to trick the FAH client into recognize GPU which is not defined in GPUs.txt?

 

I don’t want to run it long term, I just want to know if it’s possible and how it will affect PPD when folding both on host and guest.  

 

EDIT: I also wanted to mention that FAH client sees the CUDA and OpenCL device 

Don't touch GPUs.txt

If you need something changed or added, let me know.

I just need PCI ID of the device

[muziqaz]

Ah, I will not be able to add it as according to lspci output it is recognised as Microsoft device, and not Nvidia.

I can only select Nvidia, AMD or intel devices

[Kazi]

On 3/20/2025 at 11:34 AM, muziqaz said:

Ah, I will not be able to add it as according to lspci output it is recognised as Microsoft device, and not Nvidia.

I can only select Nvidia, AMD or intel devices

Yeah, unfortunately even when it is able to pick up Nvidia drivers inside that VM it still reports as Microsoft device. It makes sense because it’s a virtual device but a bit unfortunate. I am going to abandon this idea. It would probably work if I passthrough entire PCIe device however that wouldn’t help me to test if larger chips are able to process more WUs at the same time and boost overall PPD which you can get from single GPU. 

[GOTSpectrum]

4 hours ago, TheLANguy said:

The second event in PrimeGrid's 2025 challenge series, "Peter Bergmann's Birthday Challenge" will be happening from March 24 01:00 UTC to March 29 01:00 UTC. Make sure you are crunching GFN-18!

Learn more here: https://www.primegrid.com/forum_thread.php?id=10833

 

Dear LTT Folding team....