2026-09-11 — views
GPU-CFR: Compiling Game Trees for 80x Faster CFR on GPUs
Read this because This compiler approach eliminates kernel launch overhead, making GPU CFR viable for large-scale games where CPUs previously dominated.
GPU-CFR compiles game trees to static dataflow, achieving up to 80x speedup over prior GPU implementations and beating CPU baselines.
Overcoming CPU Dominance in Counterfactual Regret Minimization
Counterfactual regret minimization (CFR) has historically been a numerical workload that performs better on CPUs than on GPUs. Traditional implementations sweep game trees containing billions of states through millions of small, interdependent gather and scatter steps. Because GPU kernels execute in microseconds, the overhead of kernel launches and framework dispatch often dominates the runtime, causing prior GPU attempts to lose to optimized CPU code.
The Static Dataflow Solution
Researchers have introduced GPU-CFR, a compiler and runtime system that addresses these inefficiencies by observing that, for a fixed game, the structure of a CFR iteration is known before the first iteration begins. The system compiles any game once into a static dataflow representation. This process creates flat edge and information-set arrays, precomputed indices, and depth-level batched passes. These elements fix the entire operation sequence, ensuring that only solver state changes between iterations.
Performance Gains and Technical Innovations
The GPU-CFR architecture employs static chance folding, depth-level execution blocks, and a dual-lane reach buffer. These techniques cut the number of framework operations by up to 18.1x. Since shapes, indices, and buffer addresses remain constant, the system utilizes CUDA Graph Replay. This feature records the iteration once and replays it with a single graph launch, drastically reducing overhead.
Benchmark Results
Testing on an A100 accelerator across an eight-game suite comprising card, dice, and board games revealed significant performance improvements. GPU-CFR runs 29.8 to 80.4 times faster than the fastest prior GPU CFR implementations on the same hardware. When compared to LiteEFG, one of the fastest open-source CPU implementations, GPU-CFR is 14 to 258 times faster on the four largest games in the suite.
Notably, the compiled representation carries most of this performance margin. Even on a CPU with eight threads and no accelerator, the system is 2.2 to 51.1 times faster than the GPU baseline. On the CPU, the optimized path reproduces reference iterates bitwise, and the costs of tree construction and graph capture are recovered within the first solve. GPU-CFR outperforms every CPU and GPU baseline on the mid-to-large games of the suite without altering the underlying update rule.
Practitioner note
Developers should consider adopting static dataflow compilation for fixed-game AI solvers to eliminate kernel launch overhead. Utilizing CUDA Graph Replay can significantly reduce framework dispatch costs in iterative numerical workloads.