Building a new super fast computer

[M ANSARI]

I plan on building a new super fast computer. I still am not sure whether I am going with Intel or AMD, but that is not the real concern ... my concern is what would be best for NN engines as a GPU card. Is it better to get 2 x 2080 Ti cards or just one Titan V card? Is there anything on the horizon that makes it better to wait a little to get the correct GPU. Has anyone done some benchmarks for the NN engines on different hardware platforms?

Thanks in advance!

[Jouni]

Will You beat this?

Cluster: 128 CPUs x 32 threads = 4096 threads

2 minutes search from start position: 4.801.341.606 nodes/sec

info depth 38 seldepth 51 multipv 1 score cp 53 nodes 576165794092 nps 4801341606 hashfull 1000 tbhits 0 time 120001 pv e2e4 c7c5 g1f3 d7d6 f1b5 c8d7 b5d7 d8d7 c2c4 b8c6 b1c3 g8f6 d2d4 d7g4 d4d5 c6d4 f3d4 g4d1 e1d1 c5d4 c3b5 a8c8 b2b3 a7a6 b5d4 f6e4 d1e2 g7g6 c1e3 f8g7 a1c1 e4c5 f2f3 f7f5 h1d1 e8g8 d4c2 c5d7 a2a4 a6a5 e3d4 f5f4 d4f2 f8f7 h2h3 d7c5
bestmove e2e4 ponder c7c5

Hard times for Leela .

[Damir]

https://github.com/official-stockfish/S ... ee/cluster

[Lion]

I guess currently what makes the more sense for a high end NN system is a 2x2080Ti.
You could go with 2xTitan RTX but honestly to gain less than 25% but pay 2x the price doesn't make any sense to me.

Also what I am currently not sure is how much does Lc0 really benefit from i.e. 2x2080Ti instead of a 1x2080Ti.
I own a 2080Ti & a 2070 and I am not sure the combination is really much stronger than a single 2080Ti.
Lc0 is very fast to find a strong move but then tends to stay on the first thought....

rgds

[zullil]

Will You beat this?

Cluster: 128 CPUs x 32 threads = 4096 threads

2 minutes search from start position: 4.801.341.606 nodes/sec

info depth 38 seldepth 51 multipv 1 score cp 53 nodes 576165794092 nps 4801341606 hashfull 1000 tbhits 0 time 120001 pv e2e4 c7c5 g1f3 d7d6 f1b5 c8d7 b5d7 d8d7 c2c4 b8c6 b1c3 g8f6 d2d4 d7g4 d4d5 c6d4 f3d4 g4d1 e1d1 c5d4 c3b5 a8c8 b2b3 a7a6 b5d4 f6e4 d1e2 g7g6 c1e3 f8g7 a1c1 e4c5 f2f3 f7f5 h1d1 e8g8 d4c2 c5d7 a2a4 a6a5 e3d4 f5f4 d4f2 f8f7 h2h3 d7c5
bestmove e2e4 ponder c7c5

Hard times for Leela .

I think you are greatly overestimating the value of all those threads. What Stockfish might benefit from is the replacement of its current "lazy smp" with a high quality multi-threaded search. But that will only delay the inevitable ascension of NN chess. What Leela has already demonstrated is staggering to me; truly a new paradigm.

[ankan]

I plan on building a new super fast computer. I still am not sure whether I am going with Intel or AMD, but that is not the real concern ... my concern is what would be best for NN engines as a GPU card. Is it better to get 2 x 2080 Ti cards or just one Titan V card? Is there anything on the horizon that makes it better to wait a little to get the correct GPU. Has anyone done some benchmarks for the NN engines on different hardware platforms?

Thanks in advance!

2 x 2080Tis is close to best you can get for lc0 (it currently doesn't scale well over ~70-100 knps due to CPU side bottlenecks).
Titan V is actually slightly slower than a single 2080Ti for lc0.

If you are planning to train NNs (instead of just running lc0), higher memory capacity and higher mixed precision (fp16 multiply, fp32 accumulate) throughput of Titan cards could be useful. I would recommend the Titan RTX which has 24 GB VRAM, and full 130 TFlops of mixed precision tensor op performance. Non titan RTX series cards have full tensor throughput when doing pure fp16 tensor ops, but half throughput for mixed precision tensor ops. Running lc0 just needs pure fp16 so it doesn't matter.

Titan V is slightly slower than Titan RTX for NN stuff (fp16, mixed and fp32 throughput) and has half the memory, but it has much higher fp64 throughput so its recommended only if you have something that needs double precision (NN's don't).

Some benchmarks are here (but not directly comparable as the network ID, nncache and other setting makes huge difference in NPS):
https://docs.google.com/spreadsheets/d/ ... sp=sharing

Here are some benchmarks I did with an older version:

Code: Select all

             fp32    fp16    
GTX 1080Ti:   8996     -
Titan V:     13295   29379
RTX 2070:     8841   23721  (slightly differnt version of lc0 used, but same settings)
RTX 2080:     9708   26678
RTX 2080Ti:  12208   32472

The recently announced 2060 should be ~15% slower than 2070.

[nabildanial]

I plan on building a new super fast computer. I still am not sure whether I am going with Intel or AMD, but that is not the real concern ... my concern is what would be best for NN engines as a GPU card. Is it better to get 2 x 2080 Ti cards or just one Titan V card? Is there anything on the horizon that makes it better to wait a little to get the correct GPU. Has anyone done some benchmarks for the NN engines on different hardware platforms?

Thanks in advance!

2 x 2080Tis is close to best you can get for lc0 (it currently doesn't scale well over ~70-100 knps due to CPU side bottlenecks).
Titan V is actually slightly slower than a single 2080Ti for lc0.

If you are planning to train NNs (instead of just running lc0), higher memory capacity and higher mixed precision (fp16 multiply, fp32 accumulate) throughput of Titan cards could be useful. I would recommend the Titan RTX which has 24 GB VRAM, and full 130 TFlops of mixed precision tensor op performance. Non titan RTX series cards have full tensor throughput when doing pure fp16 tensor ops, but half throughput for mixed precision tensor ops. Running lc0 just needs pure fp16 so it doesn't matter.

Titan V is slightly slower than Titan RTX for NN stuff (fp16, mixed and fp32 throughput) and has half the memory, but it has much higher fp64 throughput so its recommended only if you have something that needs double precision (NN's don't).

Some benchmarks are here (but not directly comparable as the network ID, nncache and other setting makes huge difference in NPS):
https://docs.google.com/spreadsheets/d/ ... sp=sharing

Here are some benchmarks I did with an older version:

Code: Select all

             fp32    fp16    
GTX 1080Ti:   8996     -
Titan V:     13295   29379
RTX 2070:     8841   23721  (slightly differnt version of lc0 used, but same settings)
RTX 2080:     9708   26678
RTX 2080Ti:  12208   32472

The recently announced 2060 should be ~15% slower than 2070.

Will you be updating the list with RTX 2060 too? I'm looking an upgrade for my GTX 1060 too.

[ankan]

Will you be updating the list with RTX 2060 too? I'm looking an upgrade for my GTX 1060 too.

I don't have a 2060 to test with, but based on specs I am pretty confident that it won't be more than 20% slower than 2070 (~15% slower most likely).

[Dann Corbit]

I plan on building a new super fast computer. I still am not sure whether I am going with Intel or AMD, but that is not the real concern ... my concern is what would be best for NN engines as a GPU card. Is it better to get 2 x 2080 Ti cards or just one Titan V card? Is there anything on the horizon that makes it better to wait a little to get the correct GPU. Has anyone done some benchmarks for the NN engines on different hardware platforms?

Thanks in advance!

2 x 2080Tis is close to best you can get for lc0 (it currently doesn't scale well over ~70-100 knps due to CPU side bottlenecks).
Titan V is actually slightly slower than a single 2080Ti for lc0.

If you are planning to train NNs (instead of just running lc0), higher memory capacity and higher mixed precision (fp16 multiply, fp32 accumulate) throughput of Titan cards could be useful. I would recommend the Titan RTX which has 24 GB VRAM, and full 130 TFlops of mixed precision tensor op performance. Non titan RTX series cards have full tensor throughput when doing pure fp16 tensor ops, but half throughput for mixed precision tensor ops. Running lc0 just needs pure fp16 so it doesn't matter.

Titan V is slightly slower than Titan RTX for NN stuff (fp16, mixed and fp32 throughput) and has half the memory, but it has much higher fp64 throughput so its recommended only if you have something that needs double precision (NN's don't).

Some benchmarks are here (but not directly comparable as the network ID, nncache and other setting makes huge difference in NPS):
https://docs.google.com/spreadsheets/d/ ... sp=sharing

Here are some benchmarks I did with an older version:

Code: Select all

             fp32    fp16    
GTX 1080Ti:   8996     -
Titan V:     13295   29379
RTX 2070:     8841   23721  (slightly differnt version of lc0 used, but same settings)
RTX 2080:     9708   26678
RTX 2080Ti:  12208   32472

The recently announced 2060 should be ~15% slower than 2070.

Has someone actually tested (for instance) 4 x RTX 2070? How about 3x RTX 2070?

I guess you would have to put them in a commercial server box because I don't think you could find standard PCs with power supplies that can take that kind of a power draw.

I think that the 7 nm AMD stuff would be exciting in the 2 units x 64 core configuration. That would give 128 cores and 256 threads for chess. I guess it would blow the doors off of the GPU systems.

Big money though. I figure more than $10K just for the CPUs.

[M ANSARI]

I actually think that adding threads to an AB engine like SF reaches a point of diminishing returns very quickly. Most likely after 32 cores there is not much improvement. On the other hand, the GPU or NN engines are just in their infancy, and they are already incredibly strong. I expect some very major breakthroughs with LC0 as the people working on it get a better grip on how to utilize the hardware better. I tried a few different LC0 versions (there must be several hundreds) and it just seems that they all play totally differently from each other !!! This seems a lot like the early days of SF.