Eta chess engine project paused.

I think a parallel BestFirstMiniMax-Search with NNUE eval is worth to give a try. It would be nice to compare parallel BestFirstMiniMax with MCTS-UCT, MCTS-PUCT, MCTS-AB, MCTS-Rollouts and take a look into DAG (directed acyclic graph) but I currently lack the time to work in this. There are other chess engines out there, which combine a Best-First search (MCTS-derivative) with NNUE eval.

Some basic benchmarks of the Zeta NNUE AB framework showed 100 Knps to max 200 Knps throughput per worker for current GPUs, definitely too slow for an Eta v0501 hybrid approach considering modern CPUs with ~2 Mnps per core.

Recently the new neural network technique 'NNUE' took off on CPU based chess engines like Stockfish leveraging the vector unit of a CPU for NN inference, replacing HCE (handcrafted evaluation) with neural-networks. Hence with NNUE a hybrid design with BestFirst on CPU and MiniMax-Search with NNUE eval on GPU seems possible and in reach. The CPU-host would store and expand the game tree in memory, similar to Lc0's MCTS, the GPU would perform shallow AlpaBeta-searches (primarily quiscence-search playouts to avoid the horizon effect), similar to Lc0's MCTS-playouts.

Coupling 32 gpu-threads to one worker, assuming 2K clocks per node for move generation and AB-framework, additionally maybe 2K clocks per node for NNUE inference, results in 1.44M gpu-clocks for an 36x10 nodes q-search. In such an design the host-device-latency (aka. kernel-launch-overhead) of maybe 10 microseconds does not affect the overall performance. From entry-level GPUs with 512 cores (16 workers) to high-end-gpus with 5120 cores (160 workers) the throughput of such an parallel BestFirst on CPU and AB-playout+NNUE-eval on GPU design could range from ~11K to ~220K node-playouts/s, more than Lc0's gpu throughput but with a switch from MCTS-PUCT to parallel BestFirstMiniMax-Search and CNN to NNUE evaluation.

I am not into the details of current NNUE implementations for CPUs, therefore the estimated 2K gpu-clocks per node for NNUE inference is the biggest uncertainty.

I have no experience with running 16 to 160 parallel tasks via OpenCL on GPU, not sure if 160 unique command-queues are handable with CPU-GPU interaction.

Okay, let's do an timewarpjump back to the year 2008 and figure out how we could use the hardware back then for an neural network based chess engine.

Reinforcement Learning on a GPU-Cluster is probably a no go (the Titan supercomputer with 18,688 K20Xs went op in 2012) so we stick on Supervised Learning from a database of quality games or alike. A neural network as used in A0 with ~50 millions parameters queried by an MCTS-PUCT like search with ~80 knps is also not doable, we had only ~336 GFLOPS on an Nvidia 8800 GT back then, compared to ~108 TFLOPS on an RTX 2080 TI via Tensor Cores nowadays. So we have to skip the MCTS-PUCT part and rethink the search. Instead to go for NPS, we could build a really big CNN, but the memory back then on a GPU was only about 512 MB, so we stick on ~128 Mega parameters. So, we have to split the CNN, for example by piece count, let us use 30 distinct neural networks indexed by piece count, so we get accumulated ~3840 Mega parameters, that sounds already better. Maybe this would be already enough to skip the search part and do only a depth 1 search for NN eval. If not, we could split the CNN further, layer by layer, inferred via different waves on GPU, loaded layer-wise from disk to GPU memory via PCIe or alike and hence increase the total number of parameters...so what is the drawback if we could run an CNN with several billion parameters? Obviously the training of such an monster, not only the horse power needed to train, but the training data, the games. A0 used about 40 million RL games to reach top-notch computer chess level, for only ~50 million parameters, the Chess Base Mega Database contains ~8 million quality games...so we simply have not enough games to train such an CNN monster via Supervised Learning, we rely on Reinforcement Learning, and therefore on some kind of GPU-Cluster to play RL games... nowadays, and also back in 2008.

Eta project paused.

To run 896 gpu threads as one worker, and multiple workers in parallel, does not pay off. There is an 12x slow down when increasing from 64 to 896 threads. Actually this seems only natural, but i thought the multiple threads would handle somehow the memory latency better.

This was an attempt to use Zeta v099, a GPU AlphaBeta-search with hundreds of parallel workers, with ANNs. The overall nps throughput looked good, but the parallel AlphaBeta-search is not able to make efficient use of up to thousands of workers.

This was an attempt to use Zeta v098, a GPU BestFirst-search with thousands of parallel threads, with ANNs, but one single thread is too slow for ANN inference.

Another solution would be to perform an Best-First-MiniMax search on CPU and to do ANN evaluation on GPU. I could couple the nodes of an qsearch at leaf nodes to be evaluated in one batch to gain some nps...that's pretty much like A0 and LC0 work.

To run 1024 threads per worker will probably not work, due to register size limitation. With the OpenCL 2.x feature 'nested parallelism' it could be possible to run one thread for best-first, which calls another kernel with 64 threads for move generation and another kernel with 1024 threads for ANN inference. But current Nvidia and older AMD devices support only OpenCL 1.x, so this is not a real option.

LC0 uses batch sizes of 256 or 512 to utilize a gpu, i did a quick bench with 256 positions to be evaluated per run...

4096 nps on Nvidia GTX 750
16640 nps on AMD Fury X

Note that nn cache could double these values, but this is still far less than i could achieve when doing all computations directly on gpu device, wo host-device interaction.

And waiting for 256 positions to be evaluated at once is against the serial nature of AlphaBeta search...

One reason gpus are not used as accelerators for chess is the host-device latency.

Afaik the latencies are in the range of 5 to 10s or even 100s of microseconds, so you can get max 200K kernel calls per second per thread, even if the gpu is able to process its task much faster.

Therefore, Eta v0300, a cpu based AlphaBeta search with gpu as ANN accelerator, is flawed by design.

The overall performance of the v0400 design was satisfactory, only the nps per worker was too low, but maybe i can couple up to 1024 gpu threads to one worker, used to infere the ANN faster? Benchs will tell...

Back to cpu based AlphaBeta search with gpu ANN evaluation.

On Nvidia GTX 750 i achieve with one single cpu thread about 2 Knps, and up to ~20 Knps with 256 parallel cpu threads.

This sounds far too slow for an AlphaBeta search...

Okay, some further, not so quick n dirty, benchmarks showed

~240 nps for Nvidia GTX 750 and
~120 nps for AMD Fury X

per worker.

I assume on modern gpus about 200 nps per worker.

While NN cache could be able to double these values, this is imo a bit too slow for the intended search algorithm, considering about 36x10 qsearch positions on average per expanded node, one worker would need about a second to get a node score.

Back to pen n paper.

wip...will take some time...

* GPGPU device based
- host handles only the IO, search and ANN inference on gpu
- gpu computation will be limited by node count to about 1 second per
  repeated iteration, to avoid any system timeouts

* parallel BestFirstMiniMax-Search on gpu
- game tree in gpu memory
- best node selected via score + UCT formula (visit count based)
- AlphaBeta Q-Search performed at leafnodes to get a node score

* multiple small MLP neural networks
- about 4 million weights per network
- 30 networks in total, split by piece count

* trained via TD-leaf by pgn games
- 6/7 men EGTB could be used for training?

* 64 gpu threads are coupled to one worker
- used during move gen, move pick and ANN eval in parallel
- gpu core count depended from 64 workers to 2048 workers in total

Some quick and dirty benchmarks showed that with this design ~1 Knps per worker is possible.

Here an overview of what happened before....

Eta (0700)

* BestFirstMiniMax-Search on CPU with NNUE eval on CPU

Eta (0600)

* CNN monster with billions of parameters w/o search relies on ~billions of RL games

 Eta (0500)

* parallel BestFirstMiniMax-Search on CPU with ANN evaluation on GPU

Eta (0400)

* parallel BestFirstMiniMax-Search on GPU with ANN evaluation on GPU

Eta (0300)

* CPU based AlphaBeta search with GPU ANN eval

Eta (0200)

* fork of Zeta v099 but with neural networks

Eta (0100)

* fork of Zeta v098 but with neural networks

Since i have read the paper about NeuroChess by Sebastian Thrun i pondered on how to improve his results.

It was obvious that the compute power available in the 90s limited his approach, in training and in inference.

So he had only 120K games for training, a relative small neural network, and could test his approach only with limited search depths.

Recent results with A0 and LC0 show how Deep Learning methods profit by GPGPU, so i think the time has come to give a GPU ANN based engine a try....

--
Srdja