Stockfish on a 1995 Pentium Pro Beats 1996 Crafty on Modern Hardware 78 to 5, and Algorithms Account for 720 of the 1,335 Elo Gained Since
A modern chess engine on 1996 hardware would have beaten Garry Kasparov comfortably. That is the result of a November 2026 experiment by Nicholas Carlini, who ported Stockfish to a 200 MHz Intel Pentium Pro with 64MB of RAM and measured how much of chess engines’ strength gain since then comes from faster chips versus better code.
The Numbers
- Port: Stockfish adapted to 32-bit arithmetic and a 64MB memory ceiling, tuned for the Pentium Pro
- Estimated strength: about 3300 Elo (plus or minus 87) at classical time controls
- Reference points: Deep Blue about 2700, Kasparov just under 2800, best commodity engines of 1996 about 2300, modern Stockfish about 3700
- Total gain, 1996 commodity engine to modern Stockfish: 1,335 Elo
- Attributed to better programming: 720 Elo. The rest comes from hardware.
- Head to head: over 100 games, the modern engine on the Pentium Pro beat 1996 Crafty on a modern Ryzen 7 with 78 wins, 17 draws and 5 losses
Deep Blue was 30 P2SC processors at roughly 120 MHz plus 480 custom chess chips, analysing over 200 million positions a second. IBM spent several million dollars on it. The Pentium Pro machine would have cost four to five thousand dollars in 1995.
The Two Experiments
Updating Crafty 11.13. Carlini started from a 1996 engine. On the Pentium Pro it rates 1999 Elo at 1 second per move and 2383 at 10 seconds. Moved to a Ryzen 7, the unchanged engine hits 2614 at 1 second. Adding modern search techniques (fail-soft, late move reductions, correction history), better quiescence search and neural-net evaluation (NNUE) lifts the Pentium Pro to 2718 at 1 second and 2974 at 10 seconds. That is a 1996 engine on a 1996 chip, rated above anything that existed then.
Porting Stockfish. A naive port rates about 2496 at 1 second and 2940 at 10 seconds. After algorithm adjustments tuned for the Pentium Pro: 2846 at 1 second, 3109 at 10 seconds and 3180 at 30 seconds.
The LLM Detail
Carlini states that all the code was written by language models, a combination of Fable 5.1, Opus 5.5 and GPT-6. That includes a bootable-floppy mini operating system that puts the CPU into 32-bit protected mode, initialises a Realtek RTL8139C network card and runs an HTTP server taking FEN positions. The machine boots from a 1.44MB floppy and is playable over the internet.
The engineering here is a bare-metal port with no OS, a memory budget a modern browser tab would exceed in seconds. It is a cleaner demonstration of agent-assisted systems work than most benchmarks, with a verifiable end result: a match score.
Why It Matters Outside Chess
The 720 of 1,335 split is a data point for the argument over whether progress comes from compute or algorithms. In a domain with a hard objective measure and decades of public research, a bit more than half the gain came from software, with the remainder from hardware. Carlini’s ratings carry error bars of 34 to 97 Elo depending on the configuration, so the exact split is soft, but the sign is not in doubt.