Fast GPU, not enough CPU power to keep up?

[GarageMonster1962]

Well, I was wrong it is working but I still have the same problem. I can cotrol the cpu but not the gpu in the app. GarageMonster

[BobWilliams757]

With a GPU folding does not have any control other than working or not.

If your system has an Nvidia GPU you can simply install MSI Afterburner or several other types of software that control power limits. Often you can get a much greater deal of efficiency by reducing the max power to the GPU, but it comes at a slight cost of overall folding speed.

The software is a great way to limit power, control fan speeds, and increase the power efficiency.

[Peter_Hucker]

Anyone with overheating GPUs, try MSI Afterburner, if it's not a very old GPU there will be an adjustable temperature limiter. It just slows the clock, doesn't have to interact with folding. Or check the fans are all spinning and there's no dust. Or if you're more adventurous, remove the heatsink and put in some decent heatsink paste (I've dropped temperatures 20C by doing this, I don't know if the original stuff was rubbish or dried up).

[pcwolf]

I'm curious how this will work as I've consistently found that doing any amount of CPU Folding drags down the processing speed of the GPU job.

Its particularly aggressive on my 5950X where I will lose more PPD off the GPU job than I gain from the CPU job, and it doesn't appear to be boost clock related or down to how many cores I use. I can limit CPU Folding to one CCD, the GPU thread will be on the other CCD, but it still happens.

I don't think it has much impact on my 7800X3D system though, so I wonder if its memory latency related and the huge cache negates it?

I entered this thread to try and find an answer to that specific question myself. I jumped to AMD5/DDR5 with a Ryzen 7900 and 32 GB 6000mHz ram. Folding with the integrated Radeon GPU earns about a million and a half PPD, but causes penalty of five million PPD to the RTX 4070 Super. Clearly not worth folding with every available slot.

The entire WUs do not drive either the CPU or RAM over a quarter to a half of capacity, at least that monitors with five second intervals detect.

[Peter_Hucker]

I wouldn't have expected using onboard graphics to slow down the main graphics like that. The main graphics usually slow down if the CPU is overloaded, as that workunit needs a dedicated CPU core. What's the load on the CPU? Are you folding with that too?

[pcwolf]

ASRock B650E Taichi motherboard with one CPU PCIe 5.0 x16 slot and one chipset PCIe 4.0 x16.
First is occupied with 4070 Super and second with RTX A2000

I do not have numbers directly at hand but recall when folding with the Ryzen 7900 I also saw a penalty on the first slot card. I tried with limit to processor threads of 12 or 50% but still saw a reduction in total PPD. Maximum credit is clearly GPU with no CPU and no iGPU. Right at the moment I am logging 14.0 and 1.76 million. I am not complaining.

Curiously, I *have* run BOINC on the CPU alongside F@H on GPUs with no apparent penalty.

I would like to install beta 8 version but there is no Manjaro repository for it at the moment, and I am uncertain how to break down the Debian version to install. The pooling and distribution of WU over all available folding slot resources is intriguing. Perhaps I should just close my eyes and leap into the beta.

[Peter_Hucker]

I guess your CPU is too busy with it's own folding to service the PCI-Express lane on time. Perhaps the Boinc project you ran used a different part of the processor.

[Alex_Atkin]

I wouldn't have expected using onboard graphics to slow down the main graphics like that. The main graphics usually slow down if the CPU is overloaded, as that workunit needs a dedicated CPU core. What's the load on the CPU? Are you folding with that too?

It could be starving the CPU of memory bandwidth perhaps?

I'm seeing a core 23 WU eat 20% of a 12400 and its only GPU folding, granted on a 4090.