TalkChess.com

Hello Werner,

Werner wrote: ↑Thu Aug 27, 2020 7:36 pm I have an intel core I7 3770K Ivy Bridge.
Normally popcount versions are running here - but Igel 2.7.0 crashes (only avx).
Anybody has compiled a version running here?

Igel 2.7.0 build heavily uses AVX2 instructions and your CPU does not support it hence it crashes. I did not produce official builds other than AVX2 because older CPUs running SSE4.1 and AVX have shown performance degradation in my testing.

For people willing to use Igel for fun (like peter) I produced a private build using compile for a specific CPU architecture (westmere) but as you can understand I cannot release this version to the public because it only works good with westmere cpus.

I could produce a working solution for Ray (yet) because gcc8 I use does not have a good support for amd profiles for his cpu.

voffka wrote: ↑Thu Aug 27, 2020 10:59 pm I could produce a working solution for Ray (yet)

It is a typo,

I could not produce a working solution for Ray (yet) .

voffka wrote: ↑Thu Aug 27, 2020 11:40 pm
I could not produce a working solution for Ray (yet) .

Well, if performance on older CPUs is badly affected, then it is no use playing ratings list games on these CPUs anyway.

Thanks - not a big problem - I use my other pcs to test Igel.

Modern Times wrote: ↑Thu Aug 27, 2020 11:44 pm Well, if performance on older CPUs is badly affected, then it is no use playing ratings list games on these CPUs anyway.

It was around 40 elo less in my tests when using SSE stuff. It could be that I did not master yet compilation tricks, but it felt worrying to release it as a general availability binary.

Most of the chipsets support AVX2 since 2014, so the current binaries of Igel 2.7.0 should cover a wide range of population.

Werner wrote: ↑Thu Aug 27, 2020 11:46 pm Thanks - not a big problem - I use my other pcs to test Igel.

Great! Thank you Werner.

Guenther wrote: ↑Sat Aug 22, 2020 2:45 pm

Thanks Guenther, the avx compile is running on the older pc - but the engine is quite faster on a skylake cpu. so I think for official test I only use my skylake pc.

Hi Volodymyr !
Thank you for this new release by incorporating NNUE as a main evaluation function !
Since i own an old computer that don't support AVX2 instructions, i m also interested in having your private buid specific for westemere architecture
Thank you


[quote=voffka post_id=857858 time=1598561989 user_id=10630]
Hello Werner,

[quote=Werner post_id=857814 time=1598549817 user_id=162]
I have an intel core I7 3770K Ivy Bridge.
Normally popcount versions are running here - but Igel 2.7.0 crashes (only avx).
Anybody has compiled a version running here?
[/quote]

Igel 2.7.0 build heavily uses AVX2 instructions and your CPU does not support it hence it crashes. I did not produce official builds other than AVX2 because older CPUs running SSE4.1 and AVX have shown performance degradation in my testing.

For people willing to use Igel for fun (like peter) I produced a private build using compile for a specific CPU architecture (westmere) but as you can understand I cannot release this version to the public because it only works good with westmere cpus.

I could produce a working solution for Ray (yet) because gcc8 I use does not have a good support for amd profiles for his cpu.
[/quote]

Jamal Bubker wrote: ↑Sat Aug 29, 2020 1:14 am Hi Volodymyr !
Thank you for this new release by incorporating NNUE as a main evaluation function !
Since i own an old computer that don't support AVX2 instructions, i m also interested in having your private buid specific for westemere architecture
Thank you

OMG, westemere is getting hot Please write me in PM and I will send you the required files.

Hi Volodymyr,

I would like to use the "Night Nurse 0.2" net in fishtest for some tests in Stockfish, do you give me authorization to do so? The idea I have is to run the current Stockfish master net against the current Stockfish with Night Nurse 0.2 (I anticipate that Stockfish with Night Nurse 0.2 will be somewhat weaker, because the search is not optimized for this net), to estimate the Elo difference and maybe work on a contempt implementation for NNUE evals.

Thanks in advance,
Stéphane Nicolet