The original title is: "One RTX 5090 vs a 12-GPU Cluster — Benchmarking a Decade of GPUs on the Same Go Proof"

Short summary

A developer rebuilt a 2023 NeurIPS Go-solving system (same code, same problem) and benchmarked a single 2026 RTX 5090 against a 12-GPU GTX 1080Ti cluster from 2017. The single card was 4.64x slower in wall-clock time but needed 41% less total search work, because collocating the trainer and solver enabled 19x more frequent model updates. The benchmark cleanly isolates a decade of hardware progress and reveals that online fine-tuning payoff scales with update frequency.

• •Single RTX 5090 vs 12x GTX 1080Ti cluster on identical Go-solving task: 41h vs 9h wall-clock, but 41% less search work
• •Per-slot throughput doubled (2.01x) and search efficiency improved 1.71x due to 19x more frequent model updates from collocation
• •Clean decade-spanning hardware benchmark with honest caveats on confounded variables

Generated with AI, which can make mistakes.