Fermi Compute

[DANK_AS_gay]

I am fascinated with Fermi for no good reason, maybe because of how much of a failure it was. Something I found interesting, was the exceptional compute, which was used for, you guessed it, Folding@home. In fact, when AnandTech did a review of the 580, they actually used Folding@home as a benchmark. My question is, would it be a good idea to buy several Fermi cards (GTX580s for about $60-$80 each), and set them up in a "mining" setup but for folding@home? From what I understand, mining takes a crazy amount of memory, and memory bandwidth, but how much is Folding@home dependant on those things? Would I be able to actually contribute anything worth the effort? (energy costs aren't really an issue)

[Eigenvektor]

I found this database: https://folding.lar.systems/gpu_ppd/overall_ranks. It doesn't contain a GTX 580 unfortunately, but a GTX 570 HD. Compared to a current gen GPU it doesn't look too good.

RTX 3090    – 7,493,734 PPD
GTX 1080 Ti – 2,416,910 PPD
AMD RX 480  –   448,477 PPD
GTX 570 HD  –    63,361 PPD

You'd need roughly 118 GTX 570s to match a single RTX 3090 or 38 to match a single GTX 1080 Ti. So in terms of perf/watt or perf/$ you're much better off with a modern GPU.

 

If you can get each card at an average of $70, you'd be paying ~$8,260 to match the performance of a RTX 3090 and still $2,660 to match the performance of a 1080 Ti.

 

Of course any contribution helps. Every WU completed adds to the total number of WUs solved per hour/day/year

[Somerandomtechyboi]

Also not to mention thermi runs very hot and chugs power, better off using them as ultrabudget gaming gpus rather than mining gpus where power consumption does matter

[CWP]

Folding has a higher demand on PCIe bandwidth to keep the GPU occupied with scientific data to crunch. You cannot expect the cards to perform at their best when you have them on the 1x risers like in a mining setup.

 

Heck, I would predict if they are on the 1x risers, they might not be able to get to 100% utilization, thus not get too hot.

[damric]

F@H no longer runs on Fermi due to the updated Open CL requirement. I tried recently.

[Gorgon]

On 12/17/2021 at 1:29 PM, Eigenvektor said:

I found this database: https://folding.lar.systems/gpu_ppd/overall_ranks. It doesn't contain a GTX 580 unfortunately, but a GTX 570 HD. Compared to a current gen GPU it doesn't look too good.

RTX 3090    – 7,493,734 PPD
GTX 1080 Ti – 2,416,910 PPD
AMD RX 480  –   448,477 PPD
GTX 570 HD  –    63,361 PPD

You'd need roughly 118 GTX 570s to match a single RTX 3090 or 38 to match a single GTX 1080 Ti. So in terms of perf/watt or perf/$ you're much better off with a modern GPU.

 

If you can get each card at an average of $70, you'd be paying ~$8,260 to match the performance of a RTX 3090 and still $2,660 to match the performance of a 1080 Ti.

 

Of course any contribution helps. Every WU completed adds to the total number of WUs solved per hour/day/year

Here is an older sheet with performance figures for Fermi. So 39,000 to 43,000 PPD for a 570 or 580 or less than a 3600 CPU these days. Likely if they were supported you'd get 45 - 50kPPD with CUDA improvements.

 

These are only PCIe Gen 2 cards as well so on a PCIe2x1 riser you'd get half the bandwidth as a Gen3 so even worse.