Reported evidence
GPU-CFR → uses → Static dataflow
GPU-CFR compiles iterations for a fixed game into static dataflow, using precomputed indices and depth-level batching. Unchanging shapes, indices, and buffer addresses allow CUDA Graph Replay. This mechanism does not establish an equivalent benefit for Agents whose execution paths change with intermediate results.
Read the evidence
The GPU-CFR abstract discusses counterfactual regret minimization: for a fixed game, the structure of iterations, apart from numerical values, can be determined in advance. The authors compile it into static dataflow, precompute indices, and batch by depth; because shapes, indices, and buffer addresses remain unchanged, one iteration can be recorded and then replayed with CUDA Graph Replay.Read the full research