Folding Forum

Once upon I time I had a mining rig configured for different coins. IIRC I had set one of my 6 card rigs to fold, but reading the forums it seems like I need at least 1 logical core per GPU. To my understanding some WU's will resource lock a processor core to a job, but I also heard some months back there was an attempt to remedy this. I have a Z270 board but need a processor and RAM, I wanted to go with the good old G4400 but it's only 2 cores. I don't have a problem going and spending a little more on a i3, but the i7 is more than I want to spend. I have a lot of the parts lying around and a good deal on the GPU's but I don't want to drop a ton on the processor. The alternate idea is to go with finding a AMD mobo combo on Ebay w/ more cores and can be cheaper. Downside is finding enough PCI-E slots on said boards.

Does anyone know if an 6th or 7th gen i3 will handle the load of running 6 cards or if there was ever a patch to correct the issue of program to resource lock the your cores? I'm also told running duel channel is very helpful on some WU's as well. I don't plan on folding on the CPU so it will be free for working with the GPU's

Should mention I plan to Fold on a Linux box Using 6 1060 mining cards

The nVidia driver runs best if there's one free CPU thread per GPU. I'm not sure if the FAHClient demands it or not. The driver uses a spin-wait so the thread will appear to use 100% of the thread with no wait-states.

I have not done a lot of performance testing with AMD GPUs. The intel iGPUs and RPi 4 or above are new enough that we've received few dependable reports.

PCIe slot utilization varies somewhat with the protein assigned, but for the most part it depends on the GPU. Running a GPU on a 1x slot will work but performance will certainly be degraded. 4x or 8x v2.0 is often enough for mid-range GPUs.

And also, the number of true x16 or x8 lane slots a specific board will support may be dependent on the CPU you use, and may also factor in limitations related to using an M.2 SSD. Do your research from quality vendors websites (and/or the motherboard manufacturers site).

If slot spacing is tight, don't be afraid to use low profile GT 1030's from Gigabyte where needed.

Paul

Even on Linux FAH hogs a thread for each GPU.
The cpu doesn't just do memory swaps with the GPUs, it also does compute sanity checks on the GPUs to make sure they're behaving.
There's other threads here that talk about the bandwidth needed to feed a graphics card for FAH, IIRC it's something like PCIe 2.0 x8 or PCIe 3.0 x4 minimum for each GPU (double check, I think I'm wrong).
At the very least do some experimenting with the kit you already have for processor usage, memory usage, PCIe slot bandwidth (x1, x4, x8), power supply draw.
There will come a point where adding more cards plateaus your system score and will start to impact it (lowers the score of the whole system).
I going to suggest you do the experimenting on Windows where, I think, it's easier to benchmark and monitor your system before switching to Linux for your final FAH build.

A single PCI-E channel will carry a lot of data, I'm surprised it'll bottleneck with a 1x slot. These are only 1060's so they shouldn't hog up that much I would hope. Trying to run on 1 board is space and because I'm trying to avoid multiple systems. Multiple systems means more hardware, more heat, and more electric unfortunately. I can shove one in my server and another in my desktop and limit it to 4 in the new box maybe. I'll have to play with it some, but I was hoping to avoid ordering a CPU because an 6th or 7th gen i7 is ungodly expensive unfortunately and the i3 was much more economical.

Where would I find the PCI-E requirements for FAH anyhow?

Linux is faster than Windows for FAH and also has lower pcie requirement. For gtx 1060 you can go with pcie 3.0 x1 risers. But e.g. RTX 2080 it would bottleneck some, so for very fast GPUs you would go with pcie 3.0 x4 on Linux. Anyone tested pcie 4.0 x1 risers yet with very fast GPUs?

For 6 GPUs I would go with intel i5 6 core CPU min 2 Ghz or i7 4 core/8 threads. e.g. Intel i5 9400F

A PCIe 4.0 x1 riser is not faster than a PCIe 3.0 x1 riser, if your GPU is not a PCIe4.0 device. You would need at least a GPU of RTX3000 series or RX5000 to take benefit of PCIe 4.0.

Tuvar wrote:Where would I find the PCI-E requirements for FAH anyhow?

I don't think anyone has ever codified those requirements into any kind of simple statement. Folding will still work on a slower PCIe connection, just slower than on one that is faster. What you can find is a range of reports here and elsewhere in which people have posted their observations. Some have done tests more rigorously than others.

What reading through the various posts gives is a set of trends. Folding on the same GPU will effected more by a slower PCIe connection under Windows OS than on Linux, probably due to differences in how the OS handles the PCIe bus. Fo a high end card the difference will be greater, people have reported seeing about 10% slower for a GTX 1080(ti) class card on a PCIe 3 4x slot compared to 8x or 16x.

So for a 1060, on a 1x PCEe 3 slot you will see some decrease in folding speed compared to having run the card on 4x or faster. Linux will have a smaller decrease than Windows. If you were asking about a GT 1030 or 1050 card, you might not be able to tell the difference in speed on some projects.

Now in your post just before this you mentioned that a single channel can carry a lot of data. True, but folding on a GPU depends on very large amounts of data being transferred onto and off the GPU continually. This is not like mining where a small amount of data and calculations are transferred to a GPU for crunching.

In answer to your question whether an older gen i3 (assume quad core) could run 6 x GTX 1060 the answer is yes in both Win and Linux, would it bottle neck the answer is probably still yes. This is not a F@H requirement rather a driver and OS issue. As you are running on risers that will already cause other bottle necks you may find the overall output is similar on 4 or 6 cards and choose to run 4 for power efficiency. If I was you I might just try playing around with what your current dual core can do with 2, 3 or 4 cards running before investing too much in a quad core, but if you do have a compatible quad lying around also have a bit of a try checking output on 4, 5 and 6 cards. To do this start with all and set one at a t time to finish and see how it looks per day. There will be some outlier work units that are either higher or lower than the general average which could distort this but easy enough to jiggle active GPUs setting the one closest to finish to finish and look at impact. Please report your results here others are interested in these kind of results as still quite a number of ex mining rigs around that can potentially become folding rigs. Good luck.

It's a recurring topic on this forum.
Concerning a CPU, we've done some tests by manually throttling the CPU speed.
Turns out for GTX-like GPUs you'll need at most 2Ghz per thread.
For RTX 2000 GPUs you'll need around 1,8-2,5Ghz per thread.
RTX 3000 GPUs You'll most likely need 2,5-3Ghz CPU thread speed, with the RTX 3090 might max out around 4-4,5Ghz.
Equivalent AMD GPUs are using exactly the same CPU usage, as they pass the same amount of data per second, than on Nvidia GPUs.
The ONLY difference, is that Nivdia drivers send 'null data', which makes it appear as if your CPU is 100% active , to lock a GPU in a thread, for a slight increase in performance over not having this feat.
Without this data, they're essentially working at the same performance passing data to the GPUs.

The main core activity of say 2Ghz of actual usage for an RTX GPU is needed. THe null data can be used to share with another GPU.
For instance, if you have 4x RTX 2070 Super GPUs, and somehow your motherboard accepts 4 GPUs,
And you have a dual core 4Ghz CPU (HT disabled); your CPU should be able to share 1 core between 2 GPUs just fine (meaning without a significant performance penalty).
The benefits from gaining PPD from an extra GPU, outweigh the cons of it losing <10% of PPD on both GPUs.
For instance, if your GPUs are GT1030 GPUs, it's entirely possible to run 8-10 of these on a 4Ghz quad core, and 12-16 of these on a quadcore with HT/SMT.

Considering that most mining boards are limited to PCIE 3.0 x1 ports (some PCIE 2.0 x1 ports, these boards should not be used for folding), it is best to fold from Linux; and fold with GPUs below a 1650 Super (preferably a GTX 1060 or lower).
You could fold on RTX 2060 GPUs, but the performance per slot wouldn't be much higher than a 1060 (and would be wasted money if you ask me).
The second factor is the amount of x1 slots you have, and CPU cores and CPU speed.
If you can, use the full sized slot for the fastest GPU, if you have one. The x8 slot will more than make up for taking 8 lanes, especially if you equip that slot with a fast GPU, like an RTX 2080Ti or faster (if your motherboard BIOS can recognize the GPU).
For a 4Ghz 2 core, you're probably going to be able to run no more than 4 GPUs.
For a 4Ghz 2 core/4 thread, you're probably going to be able to run no more than 6 GPUs.
For a 4Ghz 4 core, you're probably going to be able to run no more than 8 GPUs.

Each GPU limited to the PCIE 3.0 x1 bandwidth (and under Linux).

You should also test to see if CUDA still improves folding speed, being PCIE bandwidth constrained, over doing the same via OpenCL.

I think I'm going to go with the i3 and if it's a fail I'll resell it and look at the next step. I'm only using 1060 mining cards, but I can slot 2 in a 16x slots and then add the rest as a 1x slots later as a way to check it, unfortunately because of how jobs are handed out I'm not sure how well that'll work as a solid comparison either. I'll have to tinker with it and see how it works out though. IT won't be for a week or 2 maybe since I have to order the parts from Ebay, but I'll try and remember to post the data once I get it. I have 5 pci-e slots on my server as well 3 x16 and 2 x4, it's a 1u so it'll be a odd setup it I need to go that route. That depends if they need the 6 pin power or if I can work out a way to add it since it's using server power supplies and they don't come with them heh.

I also have one of the nvidia Jetson Xavier's, will those fold?

Tinkering is good. You'll be happier with the information you gather than the collection of information you've gotten from us.

Report back so other's can see what you've gathered.

nvidia Jetson Xavier seems to have ARM-CPU and CUDA GPU. You can install the FAHclient for ARM but I think FAH will only run on X86 CPU for CUDA GPU work units? And are there any ARM-CPU work units?