The Era of Teranode per second movegeneration is here - Gigantua Warp

[mysammoore]

What I'd genuinely like to know is where you see this going as an engine, specifically what the eval does, because that's the part of the picture I can't reconstruct from the post. A perft tree is pure movegen, every child is needed and nothing a sibling returns changes what happens next, which is exactly what a warp wants. The moment there's an evaluation and a search deciding what not to look at, the node stops being a movegen problem. My engine spends about 3% of a node on movegen and about 72% on the NNUE eval, and the search is alpha-beta where a cutoff is divergence by definition, so from where I sit teranodes of movegen don't touch the thing that decides strength. So is the intent that the eval runs on the GPU as well, batched across the warp, and the search changes shape to feed it, or is the intent that this stays a perft and enumeration machine and "engine" just means the movegen core? Both are legitimate, I just can't tell which one you're building toward, and it's a very different project depending on the answer. The dedup side is the part that looks like it could carry over regardless, 65x at ply 8 is a lot more transposition than I would have guessed.

[Dann Corbit]

This is a truly amazing accomplishment. I think it can be used to produce a superior proof search chess engine, since a move generator knows if something is won, lost, drawn, or undecided.
I think it could also be used to create an incredible mate solver, since a move generator knows if something is a mate.

You may be interested in this GPU perft project:
https://github.com/ankan-ban/perft_gpu_2026

[p1x3r]

I never thought someone would be brute forcing chess this way... but GPU-based move generation looks incredibly interesting precisely because the memory model is very different and constant CPU-GPU data transfer kills performance (I guess). Anyways, I guess you'll end up doing FPGA-based move generation just to see how far you can take this.

[dangi12012]

Before branching into other topics, I would like to focus on perft(16).

I already verified perft(15) here - https://www.talkchess.com/forum3/viewtopic.php?t=86645
Which is the first independent verification. It cost around 250 usd in terms of hardware rent, but prices have risen a lot since then.

What I am very happy with is that the chess enumeration performance is really hard compute bound, and all memory shared, or global was minimized as much as possible. This means that the cheap consumer hardware is also sufficient (does not need RTX 6000 or above).

If anyone here has a way to access machines with 4x 5090 or higher equivalent, please let me know. Perft(16) is calculatable right now.