README

Research on GPU acceleration strategies for the CFR (Counterfactual Regret Minimization) poker solver.

Documents

Document	Description
GPU Acceleration Findings	Initial analysis of why naive GPU approach was slower than CPU
Batched EHS Optimization	Solution achieving 76x speedup via true batching

Key Insights

1. Kernel Launch Overhead: Sequential GPU calls (~50μs each) killed performance
2. Batching is Essential: Processing 10K+ hands per batch amortizes overhead
3. Pre-computation Works: River EHS table provides O(1) lookups during training
4. 76x Speedup Achieved: From 140 hands/sec to 10,000+ hands/sec

Related Research

• CFR Poker - Main CFR poker solver documentation