20 Graphics cards per pc

[henryzz]

https://www.techradar.com/news/you-c...ng-motherboard
Could this be of any use to anyone here?

Originally Posted by henryzz

https://www.techradar.com/news/you-c...ng-motherboard
Could this be of any use to anyone here?

https://www.asus.com/us/Motherboards...MINING-EXPERT/

[M344587487]

I've often wondered if something like this is theoretically useful for compute, which partly depends on whether your application saturates the pcie lanes. Does an LL or PRP program require much communication with the CPU in the core loop? I guess not but don't have the skills to find out. In a perfect world all data would remain on the GPU but that probably speaks volumes of my ignorance of programming on GPU.

This board seems a nice step up from the previous iteration at least, direct to USB riser is a good improvement. Linux is probably the way to go if you want 20 of the same card, windows drivers have had trouble running too many cards of the same spec in the past, maybe still (limited to 8, then 12, now I don't know).

Originally Posted by M344587487

I've often wondered if something like this is theoretically useful for compute, which partly depends on whether your application saturates the pcie lanes. Does an LL or PRP program require much communication with the CPU in the core loop? I guess not but don't have the skills to find out. In a perfect world all data would remain on the GPU but that probably speaks volumes of my ignorance of programming on GPU.

This board seems a nice step up from the previous iteration at least, direct to USB riser is a good improvement. Linux is probably the way to go if you want 20 of the same card, windows drivers have had trouble running too many cards of the same spec in the past, maybe still (limited to 8, then 12, now I don't know).

I am running the B250 Mining Expert with 10 GPUs, and of course I run 10 instances of gpuowl on it ...

[preda]

Originally Posted by M344587487

I've often wondered if something like this is theoretically useful for compute, which partly depends on whether your application saturates the pcie lanes. Does an LL or PRP program require much communication with the CPU in the core loop? I guess not but don't have the skills to find out. In a perfect world all data would remain on the GPU but that probably speaks volumes of my ignorance of programming on GPU.

Indeed, LL and PRP on the GPU require (very) little communication with the CPU. The data is not so big, and can stay on the GPU all the time. Aside from initialization, communication is needed for checkpoint save, and possibly for the PRP error-check depending on how that is implemented, but it is not a frequent check. (also a display of the current residue, if desired, requires a tiny transfer)

[kriesel]

Originally Posted by M344587487

I've often wondered if something like this is theoretically useful for compute, which partly depends on whether your application saturates the pcie lanes. Does an LL or PRP program require much communication with the CPU in the core loop? I guess not but don't have the skills to find out. In a perfect world all data would remain on the GPU but that probably speaks volumes of my ignorance of programming on GPU.

This board seems a nice step up from the previous iteration at least, direct to USB riser is a good improvement. Linux is probably the way to go if you want 20 of the same card, windows drivers have had trouble running too many cards of the same spec in the past, maybe still (limited to 8, then 12, now I don't know).

I've seen issues on Windows with 4 gpus per system; possibly thermal or power supply.

A quick look at average bus interface load in GPU-Z on CUDA gpus running a mix of TF, P-1, and LL here gives lots of 0%, a 2%, and highest was 5%, which is a gtx1070 running CUDAPm1 in a PCIe2.0 x16 slot (not 3.0 which the card supports). Bus interface load was not indicated on the AMD OpenCl cards.

[petrw1]

Let's see: 20 x 1080Ti = about 25,000 GhzDays/Day of TF or over 9,000,000 per year.

That makes me Top Producer by almost 2 to 1.

9xMillion TF's from 72-76 about 150 GhzDays per; 5.44% chance of factor.
60,000 Completions per year; 3,264 factors found.

Who wants to buy me one?
And who else wants to pay my power bill?

[CRGreathouse]

Originally Posted by petrw1

Let's see: 20 x 1080Ti = about 25,000 GhzDays/Day of TF or over 9,000,000 per year.

5000 W, and about as many dollars per year, at max consumption.

Originally Posted by CRGreathouse

5000 W, and about as many dollars per year, at max consumption.

I have measured the consumption with 10 GPUs, it is 2.2 KW/h.

[KEP]

Originally Posted by SELROC

I am running the B250 Mining Expert with 10 GPUs, and of course I run 10 instances of gpuowl on it ...

I´ve been looking at this motherboard, but it is a bit unclear for me, can this card actually support theoretically more than 1 GTX 1080 Ti card? or is it like I think I read it, that only 1 card like GTX 1080 Ti is possible to fit in this motherboard and the remaining 18 slots has to have some alternative GPU? So in short, can I fit 5 GTX 1080 Ti in this motherboard?

Regards

KEP

Originally Posted by KEP

I´ve been looking at this motherboard, but it is a bit unclear for me, can this card actually support theoretically more than 1 GTX 1080 Ti card? or is it like I think I read it, that only 1 card like GTX 1080 Ti is possible to fit in this motherboard and the remaining 18 slots has to have some alternative GPU? So in short, can I fit 5 GTX 1080 Ti in this motherboard?

Regards

KEP

The manual of B250 Mining Expert shows a figure with 11 AMD GPUs and 7 NVIDIA.