Folding Forum

Checking the forums, I didn’t see anything to directly address this. Can someone link me to a post or other resource?

I wonder which GPU, or type of GPU, would allow for maximum/best folding performance without bottlenecking.

Dell Precision T1500
16 GB RAM (maxed)
Core i5-750 (may upgrade to i7-880)
128 GB SSD w/ Ubuntu and Win10

Thanks

FAH is always bottlenecked on one or more resource. Most commonly, it saturates the GPU shaders (except if there are too many of them for the number of atoms or not enough free CPU cycles to transfer data to.from the GPU) and it always saturates the number of CPU threads that are allocated. If transfers to/from virtual memory is saturated, (known as "thrashing") then (and only then) will adding RAM actually improve performance.

plector wrote: Core i5-750 (may upgrade to i7-880)

I did some 2 Core folding along with a GTX 1050 on a i3 530 (2 Cores 4 Threads) with no problems.

I often read "use 1 core per GPU" but as i also use old hardware (some Phenom X6) its more like 1,5 cores for stable folding.
If you use a very powerful GPU I would recommend you to leave 2 cores for feeding it.


(I could give you a Xeon X3450 (4 Cores 8 Threads) if you would be in the neighbourhood)

Regards
Patrick

I've tried assigning a pair of threads to a gpu, it doesn't work. More specifically, the cores won't use any more than a single thread. The best compromise I could make was to configure the FAH client to use at least two less threads than available so the operating system has room to breath while the cores work away.

aetch wrote:I've tried assigning a pair of threads to a gpu, it doesn't work.

Its intended. Thats how a GPU slot works.

aetch wrote:The best compromise I could make was to configure the FAH client to use at least two less threads than available so the operating system has room to breath while the cores work away.

Thats what I tried to explain. Theres a need of some 'air' for the CPU to keep the system alive and to feed the GPUs.

bruce wrote:FAH is always bottlenecked on one or more resource. Most commonly, it saturates the GPU shaders (except if there are too many of them for the number of atoms or not enough free CPU cycles to transfer data to.from the GPU) and it always saturates the number of CPU threads that are allocated. If transfers to/from virtual memory is saturated, (known as "thrashing") then (and only then) will adding RAM actually improve performance.

I can understand that there’s always a bottleneck in any given system. Maybe I should have phrased it as, what is the most powerful GPU card I can use in my system that won’t be hamstrung? I think it may be a question about total work possible vs. bandwidth to move it around the motherboard. PCIe 2 16x looks like 8GB/s. Is there somewhere to check how that matches up with how much the GPU can work on?

I did some 2 Core folding along with a GTX 1050 on a i3 530 (2 Cores 4 Threads) with no problems.

I often read "use 1 core per GPU" but as i also use old hardware (some Phenom X6) its more like 1,5 cores for stable folding.
If you use a very powerful GPU I would recommend you to leave 2 cores for feeding it.

This matches up with what I’ve read here and elsewhere. It seems like, the more PCIe lanes and/or powerful the GPU, the more CPU it requires. That can mean more cores, higher clock, or both depending on what CPU and other factors. Someone here posted their best guesses for CPU need based on some CPU/GPU pairings but I can’t find it atm.

(I could give you a Xeon X3450 (4 Cores 8 Threads) if you would be in the neighbourhood)

I suspect I am not in the neighbo(u)rhood based on the spelling, but thanks anyway, that would be a cool upgrade from the i5.

Welcome to the F@H Forum plector,

I would recommend that you use Ubuntu over Windows since that would allow you to save on the PCIe overhead that happens in Windows for F@H.

RAM would be more than enough since F@H isn't RAM intensive.

This post from Meelee has some CPU frequencies matched with some GPUs: viewtopic.php?p=346440#p346440

Since PCIe 2.0 16x is almost equivalent to PCIe 3.0 8x (https://en.wikipedia.org/wiki/PCI_Expre ... _revisions) you should be able to run a wide range of GPUs depending on:
1) Your PSU size
2) Your Motherboard support - I have read that some newer GPUs would not POST on older motherboard. I think it is related to UEFI support.
3) Your Case size

plector,

As others have pointed out, your GPU options may be limited by the BIOS in that Dell. I've had HP systems where the BIOS was locked to GTX 6xx or older video cards. So that may be your limit.

That said, I have an i7-860 on an Intel DP55WB motherboard. (roughly, the same vintage as your Dell) It too has a PCIe x16 2.0 GPU socket.
I have it Folding with a Zotac GTX 1050Ti - for about 300,000 PPD with the CUDA Platform.

Home built PC
Windows 7 Pro
4GB of RAM (yikes)

I don't think I would try anything newer on my system. (RTX 2xxx or RTX 3xxx) But a GTX 750 TI, a GTX 1050 or GTX 1050 Ti should be solid Folding performers - for your older hardware.

If you can get a 1050ti to run in it, it will give you a decent number of points. You could probably get a couple more with a 3GB 1060. Power usage isnt too high for either. You may even get away with a 1650, but i dont follow the whole bottleneck threads anymore. A 1050 is doable though for sure. I used to run mine from an AMD FX AM3+ on PCIE 2.0 it gave ok PPD and used about 62w of power according to nvidia-smi.

I've got two cards on an old TA75M motherboard and A8-3870K CPU (PCIe 2 x16): GTX 1650 and RTX 2060. They return ~1.7M PPD total so they are definitely not running to full potential. I don't believe that it's the PCIe that's slowing things, as PCIe reports max 60% utilization. The startup/finish and checkpointing all take a lot of time (compared to when they were installed on a modern x570 MB w/ Ryzen 5) where the cards aren't doing anything, and the rest of the time the GPUs report 98-100%, so I feel that the CPU may be the bottleneck. This is with linux though, which people claim isn't as PCI intensive.

The main limit on your system, will be what GPU is recognized by the Bios, and what PSU you got.
I've had quite some motherboards not recognizing RTX GPUs.
And quite a few of these older motherboards don't have recent Bios updates.

In terms of performance of the CPU, you'll be limited to ~RTX 2070. Performance won't scale well beyond that.

If you had a place where you could test or return your purchases, I'd start with an RTX 3060 Ti (or 3070). Chances are very small this GPU would even be recognized by the bios, but who knows?
Maybe a Bios update enables some GPUs?
If not recognized, bump down to a 2070/2070Super (or 2060/2060 Super, or worst case a 2080 if you can get a better deal on it, since RTX GPUs are still a hard find).

Still not recognized, a 1660Ti (1660 Super, 1650Super,..). Some people have had luck with GTX 1600 series GPUs on older motherboards. Plus GPUs like these, quite often don't require an extra power cable, as they get all their power from the motherboard.
next step down a 1060 (1050 Ti, 1050), which should be an almost guaranteed compatibility for that era PCs.

If you have access to AMD, your CPU and PCIE slot will be fast enough for AMD's fastest GPU of 6 months ago (a Vega 64, or better yet, a 5700xt).
If they're not supported, you can trickle down GPUs until one is recognized and boots (5600XT, RX 580, ...).

I don't think it would be wise to invest in any GPU above a 2070/3060Ti or 5700XT, because they will be CPU bottlenecked.

Also, your PSU may be a limiting factor.
If your PC has a 500W PSU you should be fine for 1 GPU.
If it only has a 350W PSU, I would stay away from any RTX series GPU or higher, or AMD RX 570 or greater, without first upgrading your PSU.