When/why did software only engines overtake custom hardware?

Question

Back in the 1990's the top Chess computers had racks full of custom hardware to generate moves etc. These days the top chess computers all use standard CPU with no custom hardware.

What is the history of this changeover?

See also How did the engines improve since Deep Blue?

Answer 1

When did software only engines overtake custom hardware?

Pinning an exact date on this is difficult because the last great dedicated chess computer, the final 1997 version of Deep Blue, never played against software-engine-only programs.

I would guess somewhere between May 1997, when Deep Blue beat world champion Garry Kasparov 3½–2½, and November 2006 when the program Deep Fritz, running on a dual-core Intel Xeon 5160 CPU, beat world champion Vladimir Kramnik 4-2.

What is the history of this changeover?

This Wikipedia article about Deep Blue is a good source of information on this period.

Answer 2

This is more as an example than answering the question.

"Legal move generation" had dedicated hard coded hardware in systems like Deep Blue, this hardware could generate the set of legal moves from a position much faster then any CPU at the time. (Still true with FPGA.) Then a few issues hit:

• CPUs become fast enough that they could do "Legal move generation" faster than the set of legal moves could be transferred over the PCI bus, so however fast the dedicated hardware was, it stopped being useful, except on the some silicon chip as the CPU.

• The cost of creating custom CPU sized silicon chips become unaffordable for anyone working in chess computers. There were no open source CPU designs like RISC-V, so a complete CPU would need designing not just the few custom instructions.

• By the time a CPU with custom chess extensions was ready to be used, Intel would be many hardware generations ahead with commodity CPUs. So the custom CPU would need to be at least two order of multitude faster than a commodity CPU built on the same technology.

• The primary issue with large search spaces become random (impossible to cache and hard to pipeline) memory access speed, as most of the "per move code" would totally fit in the Level 0 data/instruction cache. Custom CPUs don't speed up access to very large off chip memory systems.

Answer 3

When did software only engines overtake custom hardware?

I think that your question can be answered even outside of the chess world. Historically, computers were huge machines (physically) where the hardware was so expensive that it was cheap to write the software specifically for each computer. Software was more like the filler of the blanks, rather than the focus of the product.

With the introduction of personal PCs, things changed. It was proven that cheap (standardized) hardware with generic (cheap) software might be even better than what existed previously.

And if we look at the lifetime of Deep Blue, it coincided with the shift of technologies (and mentality). It was created as a hardware-based product because that was the kind thinking in those days. It was exactly that kind of thinking that allowed IBM to consider software as irrelevant, and lose one of the biggest business opportunities of the world, possibly of all times, to Microsoft. To underline: IBM created Deep Blue ^(*) in hardware BECAUSE IBM dismissed software as unimportant.

The history after that is visible, I will not go into the details.

---

^(*) IBM was not the original creator of Deep Blue (which was originally named differently), but IBM could change the concept from hardware to software, had they a different mindset.

Answer 4

The last machine based on special hardware (FPGAs with a chess-specific configuration in this case) that was the undisputed strongest computer in the world was Hydra. Hydra was an FPGA cluster developed by a team around Christian Donninger, who was already known as a very competent chess programmer. It was probably the best chess computer in the world until around early 2006, when Rybka started to perform better than it on high-end commodity hardware.

Software is much more flexible than hardware, i.e. you can update it freely if you discover a better way of doing things. Even FPGA-defined circuits are a bit more of a headache to port to new logic (or new underlying hardware) than software running on standard hardware is. Likewise, pure software solutions profit almost automatically from the very large amount of work that hardware vendors put into improving hardware. Also, it is not clear that anything could beat current Stockfish running on good hardware at tournament time controls much of the time without pre-set openings, because chess is drawn and Stockfish is probably good enough, on good current hardware, to prove this point in practice most of the time. So it seems likely that even the best chess machine this universe allows would not have a large Elo advantage over Stockfish.

Lastly, funding for computer chess dried up quite a bit when it became clear that man vs machine matches would henceforth be uninteresting. It's possible that special hardware could still be interesting (maybe a machine with special circuitry running Stockfish's NNUE networks even more efficiently?), but nobody is willing to pay the rather significant development costs that would be necessary to prove it. This was very different in the era of Deep Blue, when special hardware was just enough to put the goal of playing at world champion level within reach.