PCIE 2.0 and CPU/GPU questions

[Flybye]

Hi all. So I have an old Socket 1155 Z68 motherboard with 6 physical PCIE 2.0 x16 slots with a 3770k CPU (4 core 8 thread). These slots will only work at x8 when more than 1 is populated which means they can handle about 4GB/s each.

I've read here each GPU should have a dedicated CPU. Is that a physical core each GPU should have or is an HT thread fine?

I was thinking about getting 3, 4, or 5 Nvidia 1660s (they exist in single slot). I'm really not sure about bandwidth requirements which is why I figured I'd finally ask on the F@H forums. What are your thoughts on having 3 or more GPUs with 4 GBs of bandwidth available to each one?

[bikeaddict]

One thread per GPU is enough.

This may be enough bandwidth for lower end cards like the GTX 1660, but I've only used PCIE 3.0 x16 with this class of hardware.

But using multiple low end GPUs is not power efficient. The 1660 will draw close to 120W and do maybe 1-1.2M PPD. A single RTX 3060 Ti or 3070 would do about 3-3.6M PPD while using about 200W, and a 4060 would probably do over 3M PPD at around 100W. The numbers on the LAR database aren't always reliable.

https://folding.lar.systems/gpu_ppd/overall_ranks

If you need inexpensive hardware with x16 slots, you can pick up used HP Z440 workstations from ebay and put a $12 Xeon CPU and cheap SATA SSD in it.

https://www.ebay.com/itm/384284924158?V ... 4284924158

[Flybye]

At first I thought of doing a bunch of single slot 1660s, but now I see they are hard to find lol. Funny thing about the bandwidth between PCIE 2, 3, and 4 is that there is a definite higher amount in bandwidth, but I've seen videos where it just doesn't make an earth shattiner difference in games. I really wonder how much of a difference does it translate with PPDs. One video I just saw took an AMD 5700XT and ran it on PCIE 2, 3 and 4. The average FPS results in 1440p in 1 game:
2.0: 90.6
3.0: 92.6
4.0: 93.2

I was really hoping to go with 4 or 5 1 slot GPUs, but you are right. It would be more practical to go with a newer faster card. I've gone through a few threads here with no definite answer to how much of a difference PCI 2, 3, or 4 make. The HP Z440 only has 2 PCIE slots, and I was hoping to have 3+ GPUs.

Right now I have a 3080, a 1050 Ti, a motherboard with PCIE 2.0, another with PCE 3.0, and another with PCIE 4.0. In the name of science, I should test both cards in each mb to finally have some real PPD numbers! It's what all curious geeks do, right?

[BobWilliams757]

The problem with testing on various hardware is that unless you get the exact work unit assignments, your testing will be skewed by the luck of the assignment servers. Even if it's the same project, some variation takes place. Though usually small, in some cases the specifics of the work units create larger variations.

Example: Running two back to back work units on project 16900 yesterday, the first delivered just shy of 1.7 million PPD, the second barely over 1.2 million PPD. This trend of variance shows across all work units done, and at various power levels.

That is a rather extreme case, more often within a project they are much closer. But any variance might mask small improvements provided by the bus speed, if they exist.


As for running multiple slower cards.... WHY? Both less cost efficient to buy as well as less cost efficient to run by far. With the money you have to spend, buying anything 4000 series is going to be the best long term bang for the buck. The 3000 series isn't much more power efficient than the Turing GPU's, just faster while still using quite a bit of power. The standard 4070 will hang with most 3080's at a little over half the power. The 4070ti is even better in power per point. I'd sell the 3080 and buy the quickest 4000 series you can stomach, watch the points fly and appreciate the long term power savings.

[Flybye]

Why? Oh, I was just trying to see how quickly to get started on a smaller budget while using old hardware. And the more I keep reading, the more I just want to forget about old GPUs. The 3080 is in my gaming rig so I'd rather not sell it or leave that pc running all the time. The f@h box I was thinking of just hiding it in an ugly frame in a cabinet that I'd remote into every once in a while. Wow, you guys are not kidding about power consumptions:

A bit different between sites, of course:

1660: 120w 1.2m
2070 super: 205w 2.4m
3050: 130w 1.1m
3060 ti: 200w 3.3m
3070 ti: 290w 4.78m
My 3080: 340w 5.4m
4060: 110w 3.27m
4060 ti: 150w 4.5m
4070: 200w 6m
4070 ti: 285w 9.88m

At this point, a 4060 Ti will probably come out cheaper than 4x1660s, probably PPD about as much, and use 150w vs 480w. I really hadn't realized how cheap the 4060s are till now. Thanks for the eye opener, guys.

[Deedss]

For what it is worth, I had an RTX 4070 on a PCIe 2.0 x 16 slot, and I had no noticable difference between that and a PCIe 4.0 x 16 system.

[Flybye]

And I've come across games with minimal difference as well. Heck, sometimes the difference is only a few frames between 2.0 and 4.0. There are plenty of benchmarks to test the bus with, but it doesn't seem to make a big difference with GPUs in the end. And f@h being so random, it's so difficult to really test the differences.