A Pentium Pro with modern chess software won 78 games against an old engine on a new PC
Nicholas Carlini’s experiment separates processor speed from improved search and evaluation, with estimated ratings clearly distinct from match results.
The old computer won the chess match. Nicholas Carlini reports that a modern engine running on a Pentium Pro beat a 1996 engine running on modern hardware by 78 wins to 5, with 17 draws. Buying the faster box did not buy the better chess player. The upgrade had been outmaneuvered.
The researcher’s experiment asks how much of chess-engine progress came from improved algorithms rather than faster processors. He used a 200 MHz Pentium Pro with 64 MB of RAM, then approached the problem from two directions: add modern techniques to Crafty 11.13, and adapt Stockfish to fit the old machine.
Carlini writes that the modern engine “wins 78, draws 17, and loses 5” in the comparison. That is his reported experiment, not a match reproduced here. Its cleanest takeaway is the directly reported 100-game result. The estimated Elo ratings elsewhere in the write-up depend on time controls and comparison methods, and the classical-time-control figure is extrapolated. A number wearing a plus-or-minus sign still needs its method attached; the punctuation is not a tiny lab coat.
Modern Stockfish could not simply arrive on the Pentium unchanged. The port needed to work with 32-bit arithmetic, fit within the available memory and avoid assumptions about modern hardware. Carlini describes shrinking its neural evaluation, removing features such as threading and tablebases, and tuning it for shallower searches. Moving an engine backward in time is quite a bit more involved than selecting a retro desktop theme.
The Crafty path asked the inverse question. Starting from an old engine, changes to search, evaluation and tuning improved play on the old processor. That is useful because it lets the account discuss concrete mechanisms instead of handing all improvement to a vague notion of smarter software. Better choices about what to examine can help even when the machine has not become faster.
The write-up includes individual steps whose measured gains differ at one second per move and ten seconds per move. Some changes help speed more; others become useful when deeper search can exploit a better evaluation. A single total score hides those tradeoffs. The clock matters to the algorithm as well as to the player staring nervously at the board.
The author also flags a limit in comparisons to human grandmasters. Ratings from different ladders are not perfectly interchangeable. His claim that the machine would likely have beaten Kasparov is therefore an inference, not a newly discovered historical match. Kasparov did not secretly travel forward in time to validate a benchmark. That would require a rather different set of notes.
The practical lesson for someone choosing compute is to evaluate the actual software doing the job. Old code on a new processor and improved code on old hardware can be very different propositions. That does not make obsolete equipment a universal bargain; it makes an algorithm and its workload part of the purchasing question.
A follow-up could apply the same discipline to another well-defined task with repeatable inputs and clear outcomes. The unanswered question is how much comparable software progress remains available in that task. This chessboard has already supplied a concrete result, and the Pentium is still sitting on the winning side.
Pentium takes the new machine’s lunch.