GPU-CFR: 80x Faster Counterfactual Regret Minimization by Compiling the Game to Static Dataflow and CUDA Graph Replay
GPU-CFR compiles a fixed game into static dataflow with flat edge and information-set arrays, precomputed indices, and depth-level batched passes. Static chance folding, depth-level execution blocks, and a dual-lane reach buffer reduce framework operations by up to 18.1x. CUDA Graph Replay records the iteration once and replays it with a single graph launch. On one A100, GPU-CFR achieves 80x faster CFR compared to prior GPU implementations.
Counterfactual regret minimization (CFR) is a large numerical workload that has historically run faster on CPUs than GPUs due to millions of small interdependent gather/scatter steps and kernel launch overhead. GPU-CFR exploits the fact that for a fixed game, the entire operation sequence except numerical values is known before the first iteration. It compiles the game once into static dataflow, reducing framework operations by up to 18.1x, and uses CUDA Graph Replay to launch the entire iteration as a single graph. On one A100, this yields an 80x speedup over previous GPU CFR implementations.
The key technical advance is treating CFR as a static dataflow problem for a fixed game, enabling aggressive compile-time optimization and elimination of per-kernel launch overhead via CUDA Graph Replay. The dual-lane reach buffer and depth-level execution blocks suggest careful memory access pattern optimization for irregular tree traversals on GPUs.
This work could make GPU-accelerated CFR practical for large imperfect-information games, potentially reducing training time and cost for game-solving AI systems. It may also inspire similar static compilation approaches for other iterative graph algorithms on GPUs.
Faster CFR on GPUs could lower the computational cost of training game-solving AI, making it more accessible for research and commercial applications in strategic decision-making, auctions, and security games.
Next signals to watch include whether GPU-CFR is applied to larger games like full-scale poker or other imperfect-information settings, and whether the static dataflow compilation approach is generalized to other iterative numerical workloads beyond CFR.