The Pentium Pro Beat the Supercomputer on Software

The Pentium Pro beat the supercomputer on software: Carlini's Stockfish port measured 3180 Elo, and most of thirty years of chess gains came from code.

The Pentium Pro beat the supercomputer on software. Nicholas Carlini ported Stockfish to a 200 MHz Pentium Pro with 64 MB of RAM, measured 3180 plus or minus 80 Elo at 30 seconds per move, and attributed +720 of +1335 Elo since 1996 to better programming, not faster chips. Kasparov was just under 2800, so the beige box wins.

In 1997 IBM beat Garry Kasparov with Deep Blue: 30 processors, 480 custom chess chips, and according to Carlini, the 259th-ranked supercomputer in the world at the time.

This month Carlini answered the question nobody asked out loud: did they need the supercomputer? He bought a used 200 MHz Pentium Pro with 64 MB of RAM, ported Stockfish onto it, tuned it for that chip, and measured it at 3180 plus or minus 80 Elo at 30 seconds per move. Kasparov was just under 2800.

The beige box wins. Easily.

Could a modern chess engine on a Pentium Pro beat Kasparov?

Carlini’s setup is a period box, not a museum piece with a modern board stuffed in the case. 200 MHz Intel Pentium Pro on an Intel VS440FX motherboard, 64 MB of SIMM RAM at 66 MHz, booting from a 1.44 MB floppy, talking to the network through a Realtek RTL8139C. Serial console to a modern desktop. He dates the Pentium Pro to November 1, 1995, thirty-one years before the post.

He puts a price on it so you don’t confuse it with Deep Blue. The CPU sold for $1,325 at launch. Four 16 MB SIMMs would have been about $2,000. Whole machine, maybe four or five thousand dollars: a nice workstation, not a supercomputer. By 1997, when the rematch happened, he estimates the same box might have gone for about a thousand dollars.

Deep Blue, the upgraded 1997 machine, used 30 P2SC processors at roughly 120 MHz and 480 custom VLSI chess co-processors for move generation and evaluation. Carlini says it could analyze over 200 million chess positions per second. The first time a computer beat Kasparov in a game was February 10, 1996. Deep Blue lost that match and won the rematch 3.5-2.5.

Carlini’s question is narrower: could commodity 1996 hardware running a smarter engine have done the job? He says yes, and it almost certainly would not have been close.

Skill in chess is Elo. He reminds you what the gaps mean: 200 Elo is about three wins in four, 400 Elo is about a 90% expected score, 1000 Elo is about 99.7%. Best engines on commodity 1996 hardware were roughly 2300. Deep Blue was about 2700. Modern Stockfish is closer to 3700.

The Pentium port’s measured 3180 at 30 seconds per move sits well above Kasparov’s just-under-2800. The ~3300 classical estimate is higher still. Carlini is careful about ladder differences between FIDE and computer lists. He thinks a 3200-3300 engine versus a sub-2800 human still has enough buffer. He also says: if a 2700-plus super-GM wants to try, the box is on the internet.

How much of the Elo gain is software versus hardware?

His headline split is the one I keep thinking about. Of the +1335 Elo gained since 1996, he attributes +720 to better programming, not faster hardware. More than half. Search algorithms, tuning, a small neural net eval, and a pile of patient fixes, all running on a machine that boots from a floppy.

He ran it both directions.

Old engine on new hardware: Crafty 11.13 from 1996 goes from 1999 plus or minus 90 on the Pentium Pro to 2614 plus or minus 97 on a Ryzen 7, both at one second per move. At ten seconds on the Pentium, the 1996 Crafty is 2383 plus or minus 77.

New ideas on old hardware: he backported modern search and NNUE into Crafty and got it to 2718 plus or minus 34 on the Pentium Pro at one second per move, and 2974 plus or minus 74 at ten seconds. Same box, same clock, better method. The steps in that table include fail-soft, late-move reductions, SPSA tuning, Texel-tuned eval, a small net trained on self-play, then a stronger teacher, then reverse futility pruning and a wider net. Some per-step gains are smaller than their error bars. He prints those too.

Then the inverse: Stockfish 17 squeezed onto the Pentium. Naive port, 32-bit arithmetic, no SIMD, no 64-bit registers, no C++ standard library, no threads, no tablebases, a much smaller NNUE than the 71 MB original: 2496 plus or minus 64 Elo at one second, 2940 plus or minus 76 at ten. After tuning the algorithm to a slow chip (cheaper eval, on-policy NNUE, SPSA for search constants that assumed millions of nodes per second, correction history, compiler flags and 32-byte alignment): 2846 plus or minus 35 at one second, 3109 plus or minus 79 at ten, 3180 plus or minus 80 at 30 seconds.

And the tidy summary: a 100 game match between modern-engine-on-Pentium-Pro and 1996-Crafty-on-Ryzen. 78 wins, 17 draws, 5 losses for the old hardware.

Three ways to get stronger at chess, in his framing: analyze more moves, analyze better moves, or analyze moves better. Humans mostly do the last two. Engines dominate the first. Most recent engine progress, he says, looks like the human path: better move selection and better evaluation, because there is not that much room left to make a given CPU see more positions per second.

Why do Carlini’s error bars matter?

Here’s the part I’d frame and hang in a test lab. His big headline number is about 3300. He tells you right next to it that it’s extrapolated, because he can’t run enough classical games on one machine. The measured number is 3180 at 30 seconds per move, and he shows how he got from one to the other (about +70 Elo per tripling of time, then a haircut for diminishing returns: six minutes per move in a classical game would naively suggest about 3350, and he calls 3300 a safe estimate). Every figure in his tables has an error bar. Some of the per-step gains are smaller than their own error bars, and he says so.

That’s how a measurement is supposed to look. I made the same argument in One Run Is an Anecdote, and Carlini just did it better with a Pentium. Report the measured number, the time control, the uncertainty, and the extrapolation method. Do not lead with the rounded marketing figure and hide the footnote.

A worked example from his own arithmetic, so the traveler is copyable:

  1. Write down the measured point: 3180 plus or minus 80 Elo at 30 seconds per move.
  2. Write down the scaling you used: about +70 Elo per tripling of time between 1s, 10s, and 30s.
  3. Write down the target time control and the haircut: classical is longer, diminishing returns apply, so 3300 not 3350.
  4. Write down what you did not measure: enough classical games on one machine.
  5. Keep failed or noisy steps in the table. If a change is smaller than its error bar, it stays in the table as that.

If your agent benchmark does not survive that list, it is a story about one run.

What does this mean before you buy another GPU?

Every test engineer learns this one early. When a measurement is noisy or slow, the first instinct is to buy a better instrument. Most of the time the fix is the method: better fixturing, shorter leads, a smarter sequence, averaging done right. The new scope is great. The new scope on a bad method is an expensive way to get the same wrong answer faster.

If you run models at home, this is the useful takeaway. Before buying the next GPU, look at what software tuning gets you on the hardware you have: quantization choice, context size, batching, the runtime itself. I already made the house version of that argument in run the model on the quiet chip: energy and heat are the spec for a box that sits in a living room all day. Method first, then the next watt.

Hardware is AI’s teeth is still true. You need the hall, the power, the robots assembling the racks. Carlini is not saying hardware is irrelevant. He is saying the first time we solve something, we usually solve it with brute force, and the better method shows up later. Deep Blue was the brute-force station. The Pentium running Stockfish is the later method on the cheaper fixture.

The dedicated versus fungible chip fight is the same floor argument at datacenter scale: do you buy a special-purpose chess chip, or do you make the general station smarter? Carlini’s result is one data point on that fight, for a closed game, with error bars. It is not a siting permit and it is not a GPU shopping list.

Carlini is careful here and I’ll be careful too. He mentions the obvious question, whether software gains in language models will outpace hardware scale, and says plainly that his toy experiment doesn’t answer it. It “gestures at a potential future.” That’s all. Chess is a closed game with a clean score. Don’t turn this into a forecast about AI progress. He didn’t.

He also notes a meta-lesson he is willing to own: he asked language models to write most of the engine-port code, and he wrote the article. That is a workflow note, not a claim that LLMs will follow the Elo curve.

What the experiment does say is narrower and more useful: the first time we solve something, we usually solve it with brute force, and the better method shows up later. If you’re about to spend money on more hardware, spend an afternoon on the method first. The floppy disk might be enough.

Frequently asked questions

Could a modern chess engine on a Pentium Pro beat Kasparov?

Nicholas Carlini’s port of Stockfish to a 200 MHz Pentium Pro with 64 MB of RAM measured 3180 plus or minus 80 Elo at 30 seconds per move. He estimates about 3300 at classical time controls, extrapolated from shorter games. Kasparov was rated just under 2800. Carlini calls a win overwhelmingly likely, with enough buffer to absorb FIDE versus computer-list differences.

How much of chess engine progress came from software rather than hardware?

By Carlini’s measurement, +720 of the +1335 Elo gained since 1996 is attributable to better programming rather than faster hardware. A 100-game match of modern-engine-on-Pentium-Pro versus 1996-Crafty-on-Ryzen finished 78 wins, 17 draws, and 5 losses for the old hardware.

What was the measured number versus the headline number?

The measured number is 3180 plus or minus 80 Elo at 30 seconds per move. The headline of about 3300 is extrapolated: about +70 Elo per tripling of time from 1 second to 10 to 30, then a haircut for diminishing returns toward classical time controls. Carlini says he cannot run enough classical games on one machine.

Does this prove language models will improve more from software than from hardware?

No. Carlini says the toy experiment does not answer that question. It gestures at a potential future. Chess is a closed game with a clean score. Do not turn the Pentium result into a forecast about AI progress.

Disagree or have a war story? Reply on X