CSTal2.07-Scalar builds for Legacy PCs now released

[chrisw]

For those people who've been asking for several years for a Chess System Tal that runs old old legacy PC's, at last, finally, a scalar legacy build that computes identical evals, search trees, and PVs to the AVX2 release — even the floating-point pruning nets produce the same decisions. The only difference is speed: AVX2 runs (my laptop) ~1.5M nps vs scalar ~350k nps (≈4–5× slower), which is the expected, acceptable trade-off for "runs on any x86-64 CPU."

Available on https://whittingtonchess.com/

[kaissa]

Thank you very much for this lovely gift.

[Jjaw]

Thanks so much for these versions. I am old and retired on a modest income and can't afford avx laptops.

[chrisw]

Thanks so much for these versions. I am old and retired on a modest income and can't afford avx laptops.

Pleased to help, sorry it took so long. Let us know how you get on, the 4x speed reduction is much, and it will be compounded by reduced functionality of a legacy PC

[Jjaw]

It didn't work. When I tried to load the engine it closed and produced an error window that said it could not find an entry point into some kernel.

[chrisw]

It didn't work. When I tried to load the engine it closed and produced an error window that said it could not find an entry point into some kernel.

I lack a legacy PC to test against. Could be that the footprint is just too much for an old PC, but if you can send as much failure detail as you can, there might be some information useable

[chrisw]

It didn't work. When I tried to load the engine it closed and produced an error window that said it could not find an entry point into some kernel.

Try downloading again, I forget to tell the compiler not to optimise with SIMD, now there should by no advanced instructions in the code to crash your system.

[pohl4711]

For those people who've been asking for several years for a Chess System Tal that runs old old legacy PC's, at last, finally, a scalar legacy build that computes identical evals, search trees, and PVs to the AVX2 release — even the floating-point pruning nets produce the same decisions. The only difference is speed: AVX2 runs (my laptop) ~1.5M nps vs scalar ~350k nps (≈4–5× slower), which is the expected, acceptable trade-off for "runs on any x86-64 CPU."

Available on https://whittingtonchess.com/

There is a problem with the Absurd avx512 binary: Threads are limited to 32 instead of 64?
And (more strange): When I set Hash to 256, as it is default in the avx2 version (why the number is different in avx512 binary??), I still get a different eval and pv-line by both binaries, if I do a "go depth 25" in console mode. Which should not happen, if the same sourcecode is compiled in avx2 and avx512. So, there is definitly something broken here.

[chrisw]

For those people who've been asking for several years for a Chess System Tal that runs old old legacy PC's, at last, finally, a scalar legacy build that computes identical evals, search trees, and PVs to the AVX2 release — even the floating-point pruning nets produce the same decisions. The only difference is speed: AVX2 runs (my laptop) ~1.5M nps vs scalar ~350k nps (≈4–5× slower), which is the expected, acceptable trade-off for "runs on any x86-64 CPU."

Available on https://whittingtonchess.com/

There is a problem with the Absurd avx512 binary: Threads are limited to 32 instead of 64?
And (more strange): When I set Hash to 256, as it is default in the avx2 version (why the number is different in avx512 binary??), I still get a different eval and pv-line by both binaries, if I do a "go depth 25" in console mode. Which should not happen, if the same sourcecode is compiled in avx2 and avx512. So, there is definitly something broken here.

I’ve researched the threads/cores issues with people who know what they talking about. Setting threads greater than cores has two negative effects, it puts heavy pressure on the cpu, especially when AVX2 or AVX512 SIMD heavy (as is a neural net engine) which gives thermal problems and throttling problems. Dependent on the cooling system, high thermal loads can shut down your system and you get a grey screen crash. Secondly, if you make threads more than cores, the system still only has limited CPUs and the multi engines will be cycled round the available cores, this slows things down down pretty much if not worse than the attempt being made to get too many threads. So, thread limit = number of cores. You may not like it but you’re wrong. The problem is your false expectations of gaining from threads more than cores.

Default hash has been changed between recent versions. As has minimum hash. Again all with intention of preventing users from setting insanely high thread counts (which immediately impact engine footprint) and running OTT concurrencies. I would guess you are using differently numbered release versions if you get differing AVX2/512 behaviour.

Last thing, some couple of weeks ago I emailed you that I had produced a bit identical EAS profiler to the one you release (yours in batch file language), that mine was 3x faster and also removed the bug I found in your code. I asked you if you were interested in having 3x, big free and maintainable (because not written in terrible batch file language) and to get in touch if so. I have not had the courtesy of a reply. Presumably you are happy with slow, badly written, unmaintainable and buggy test software being used by all and sundry and have no interest in a proper version? Fine, no problem, but be aware that people in glass houses should avoid throwing stones.

[tapio]

I say thank you - will eagerly, but in a few days, elders moved in an everything changes.