Page 1 of 1

rtx 2080ti scaling

Posted: Tue Feb 04, 2020 10:09 pm
by stavros
hi.has anyone some data about 2080ti scaling for 1,x2,x4 sli 2080ti only for chess and Lc0? tnx

Re: rtx 2080ti scaling

Posted: Tue Feb 04, 2020 10:39 pm
by brianr
stavros wrote:
Tue Feb 04, 2020 10:09 pm
hi.has anyone some data about 2080ti scaling for 1,x2,x4 sli 2080ti only for chess and Lc0? tnx
I am far from an expert (and only have one 2080ti), but SLI does not help (and may slow down) Lc0.
For some benchmark info see:
https://lc0.org/benchmarks

Re: rtx 2080ti scaling

Posted: Wed Feb 05, 2020 11:14 am
by crem
brianr wrote:
Tue Feb 04, 2020 10:39 pm
stavros wrote:
Tue Feb 04, 2020 10:09 pm
hi.has anyone some data about 2080ti scaling for 1,x2,x4 sli 2080ti only for chess and Lc0? tnx
I am far from an expert (and only have one 2080ti), but SLI does not help (and may slow down) Lc0.
For some benchmark info see:
https://lc0.org/benchmarks
Just to clarify, having multiple GPUs do help, it's just that they don't need to be connected with SLI.

Re: rtx 2080ti scaling

Posted: Wed Feb 05, 2020 10:52 pm
by jjoshua2
The rule of thumb is that scaling is perfectly linear as expected adding cards, until you get to around 50k nps, and then going downhill to basically zero scaling once it hits 100k nps (2 to 4 thread CPU speed is what determines these numbers, IPC, boost, memory latency). Advertised max CPU boost speeds often aren't hit with 3 or 4 threads.

So how many cards depends on how fast each GPU gets and what size net you can do. For example a 10x128 might already be about maxed on a $300 RTX card and a 30x384 might not be maxxed on $32,000 4 x V100. And 6 to 8 GTX 1060s apparently scales quite well with most net sizes.

Re: rtx 2080ti scaling

Posted: Thu Feb 06, 2020 6:02 am
by Laskos
jjoshua2 wrote:
Wed Feb 05, 2020 10:52 pm
The rule of thumb is that scaling is perfectly linear as expected adding cards, until you get to around 50k nps, and then going downhill to basically zero scaling once it hits 100k nps (2 to 4 thread CPU speed is what determines these numbers, IPC, boost, memory latency). Advertised max CPU boost speeds often aren't hit with 3 or 4 threads.

So how many cards depends on how fast each GPU gets and what size net you can do. For example a 10x128 might already be about maxed on a $300 RTX card and a 30x384 might not be maxxed on $32,000 4 x V100. And 6 to 8 GTX 1060s apparently scales quite well with most net sizes.
You mean NPS scaling? How about strength scaling 1 -> 2 -> 4 -> 8 when the NPS is not maxed (say a large net on RTX 2060 GPU)? Wouldn't there be a larger and larger proportion of wasted nodes, even if NPS scales perfectly?