GPU-CFR Compiles the Game to Static Dataflow and Beats the Fastest Prior GPU CFR by 29.8-80.4x
Counterfactual regret minimization has been one of the few big numerical workloads that runs faster on CPUs, because each iteration issues millions of tiny interdependent gather and scatter steps and kernel launch plus framework dispatch dominates GPU time. GPU-CFR exploits the fact that for a fixed game everything about an iteration except the numbers is known in advance, compiling the game once into flat edge and information-set arrays with precomputed indices and depth-level batched passes, cutting framework operations up to 18.1x, then recording the iteration with CUDA Graph Replay and relaunching it as a single graph. On one A100 across eight games it runs 29.8-80.4x faster than the fastest prior GPU CFR and 14-258x faster than LiteEFG on the four largest games, with the compiled representation alone worth 2.2-51.1x on eight CPU threads.
↳ Follow the thread