Add a pcie 1x card?

[ProDigit]

My server is almost set up.
It has a GTX 1050 OC 2GB in the PCIE 16x slot.

Would my computer fold any better if I added a $40 GeForce GT 710 graphics card to the mix?

I know people are saying the PCIE 1x slots are too slow, but how slow are they, compared to a CPU?

The GeForce GT 710 is the fastest pcie 1x card available.
Or would you recommend me to go cheaper, and get a lower performing, lower price card instead? (Maybe because the 710 would never fully load under a pcie1 slot, and it would make more sense to get a card that would be under full load at pcie 1x speeds?)

Or would you say that the client doesn't recognize 2 Totally different graphics cards?

Or perhaps the slower card might slow down the faster card?

I'd be interested to know if cuda works with a fast, and slow card from NVidia.

Any answers are greatly appreciated!
Thanks!

[Asgaroth]

Disclaimer: I've only been folding for about a week and a half now, so, more experienced folders responses should take preference.

My experience is thus:

I have 6 GTX 1070 TI's, all of them plugged into a x1 PCIe slot each (Currently running at PCIe Gen-2 speeds (5GT/s). These cards are near fully "loaded" in terms of GPU core usage, however, compared to the expected speed from a GTX 1070 TI as referenced in the documented speeds google document, these are currently bouncing between 680K - 710K PPD each, which equates to about a 10 - 20% performance drop when compared to running at PCIe link width x8 or x16 at Gen-3 speeds.

Load on GPU's

Code: Select all

Tue Dec 18 17:02:15 2018       
+-----------------------------------------------------------------------------+
| NVIDIA-SMI 415.22       Driver Version: 415.22       CUDA Version: 10.0     |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|===============================+======================+======================|
|   0  GeForce GTX 107...  Off  | 00000000:01:00.0 Off |                  N/A |
| 41%   55C    P2   158W / 180W |    139MiB /  8119MiB |     97%      Default |
+-------------------------------+----------------------+----------------------+
|   1  GeForce GTX 107...  Off  | 00000000:0B:00.0 Off |                  N/A |
| 43%   56C    P2   142W / 180W |    163MiB /  8119MiB |    100%      Default |
+-------------------------------+----------------------+----------------------+
|   2  GeForce GTX 107...  Off  | 00000000:0D:00.0 Off |                  N/A |
| 46%   60C    P2   168W / 180W |    155MiB /  8119MiB |     97%      Default |
+-------------------------------+----------------------+----------------------+
|   3  GeForce GTX 107...  Off  | 00000000:0E:00.0 Off |                  N/A |
| 43%   57C    P2   130W / 180W |    139MiB /  8119MiB |     97%      Default |
+-------------------------------+----------------------+----------------------+
|   4  GeForce GTX 107...  Off  | 00000000:0F:00.0 Off |                  N/A |
| 46%   60C    P2   148W / 180W |    155MiB /  8119MiB |     97%      Default |
+-------------------------------+----------------------+----------------------+
|   5  GeForce GTX 107...  Off  | 00000000:11:00.0 Off |                  N/A |
| 45%   59C    P2   173W / 180W |    163MiB /  8119MiB |    100%      Default |
+-------------------------------+----------------------+----------------------+
                                                                               
+-----------------------------------------------------------------------------+
| Processes:                                                       GPU Memory |
|  GPU       PID   Type   Process name                             Usage      |
|=============================================================================|
|    0     25718      C   ...D64/NVIDIA/Fermi/Core_21.fah/FahCore_21   129MiB |
|    1     25561      C   ...D64/NVIDIA/Fermi/Core_21.fah/FahCore_21   153MiB |
|    2     25530      C   ...D64/NVIDIA/Fermi/Core_21.fah/FahCore_21   145MiB |
|    3     25688      C   ...D64/NVIDIA/Fermi/Core_21.fah/FahCore_21   129MiB |
|    4     25651      C   ...D64/NVIDIA/Fermi/Core_21.fah/FahCore_21   145MiB |
|    5     25625      C   ...D64/NVIDIA/Fermi/Core_21.fah/FahCore_21   153MiB |
+-----------------------------------------------------------------------------+

PCIe link width and speed (only referenced one card here, but all are the same:

Code: Select all

                LnkSta: Speed 5GT/s (ok), Width x1 (downgraded)
                        TrErr- Train- SlotClk+ DLActive- BWMgmt- ABWMgmt-

For me this is great, I'm not looking to be the #1 folder, I'm just looking to contribute as efficiently as I can.

I guess the question you need to ask yourself is, what slots do you have available, can your CPU drive the slots/gpu's available, if so, go for it, the way I see it is, every little bit helps.

As for comparisons against a CPU, I cannot say, I've not tested it, however, I would expect a GPU to be quicker based on the data it is crunching.

EDIT: As for the CUDA question, my understanding is that the current FaH client is not using the CUDA framework, it uses the OpenCL framework.

[bruce]

I have a machine with a GT710 and an (old) dual AMD CPU. The real question should be to compare the performance of it if it's configured to fold one WU with 2 CPUs or 2 WUs with a single CPU a single GPU. (Remember the GPU will reserve one of my CPUs to support it. (I have not benchmarked those two configurations.)

Yesterday, my single CPU was running a WU that was predicted to be finished in ~15 hours and the timeout for that WU was ~15.6 hours. That means almost anything that happened to delay that WU would be costly in terms of the (extremely small) bonus points it might earn -- and from a scientific standpoint, I could easily waste the entire processing because the WU would be declared "lost" and would be reissued to someone else, even if I turned it in late.

A single GT710 would earn slightly more than a CPU, but a single GT710 MINUS a single CPU is of such a small contribution that I wouldn't waste any time worrying about it.

[ProDigit]

Interesting!

I should have 10 cores, 20 threads available in my Xeon processor, running at between 1,6 to 2Ghz; rather slow; but enough cores.
The PCIE 16x slot has a GTX 1050 in it.

I only have 2 more PCIE1x slots available,

1 of them could host a secondary graphics card, but the other not, because there's no space in the back or sides to fit one.

Are there some sort of graphics accelerator cards available that can work in a narrow PCIE 1x slot?

I'm also interested in if matching 2 different graphics cards work via CUDA, or if one needs the same graphics card?

From what I'm reading, if the GTX 1070 is running at nearly 80% of the PPD speed, I could add a 1050 in the secondary slot, as it is a cheap card to get.
I only don't know if the PCIE 1x slot will have enough power, hence why the GT 710.

[Asgaroth]

If you really want to use the PCIe x1 slot(s), you could look into getting powered risers for them. These are the risers I'm currently using in my system, currently sold out.

If you do want to use powered risers then look for the 009S versions of them, previous editions could be a little hit n miss. (alternatively the 007C versions were good too).

You would still need to find somewhere to mount the gpu's though, if you dont have space in your case to mount them then I guess your only choice is the GT-710, but, take heed to what bruce mentions above about that card.

[ProDigit]

I know, but I have a Xeon processor, running at 1,6Ghz.
The cores are pretty slow themselves, but I have plenty of them.

I just read that my GTX 1050 doesn't support SLI, thus doesn't support Cuda.
Does the Folding client recognize multiple cards of different make, even if they're not configured in SLI?

[Joe_H]

Supporting CUDA and SLI are two unrelated issues. As far as I know there are no current nVidia cards that do not support CUDA?

In any case, the F@h GPU cores use OpenCL whether the card is from AMD or nVidia.

[foldy]

FAH does not use SLI but calls each GPU individually - so mixed GPUs are supported.

[bruce]

I know, but I have a Xeon processor, running at 1,6Ghz.
The cores are pretty slow themselves, but I have plenty of them.

You missed my point entirely. The fact that you have "plenty" CPU cores and I was talking about my system which only has two doesn't change the fast that every time you add a GPU (including a GT710 or something else) you will reduce the number of free CPU cores by one, so adding the PPD of a GPU also involves reducing the PPD of any CPU slot that's folding. [Plenty minus one} is less than [plenty]so [plenty plus a GPU] adds less than simply adding the GPU.

[ProDigit]

Ok, now I'm missing the point.
Low power hyperthreading cores perform rather poorly on fah client. Much poorer than a corei3 or i5.
Replacing a single 1.6-2Ghz core, with 192 stream processors operating at 954Mhz, looks like a good deal to me!
I presume adding just any graphics card, so long it's supported, will work better?

[bruce]

Yes, adding 192 available stream processors and reducing your available system by one CPU Ht thread will be a significant net benefit.