A version close to SF 11 was around 3035 at 20 kn/s.
A Pentium 75 would run SF 11 at speed around 6 kn/s. Rating for SF 11 on this P75 would be around 2900-2950.
SF 12 is around 50 RP over SF 11.
I'm afraid the network file may not fit on the harddisk of the i486.
In late 1993 early 1994 I built a raid array running on a Netware server. It used a CMD CRD-5000 Raid controller. The controller could support 7 strings of 4 drives each. Each had an 80-pin SCSI SCA connector. I'm not 100% sure but, I think the drives were low voltage differential type. The original plan was to attach 7 x 4-Gb Seagate 7200RPM Barracuda drives to it. Due to supply problems I ended up using Quantum drives. This gave 28Gb of raw storage, and a little less than 24Gb in a RAID 5 configuration. Eventually this system supported over 70GB of disk storage. This server supported SGI, some proprietary Unix and windows system.
My home system at the time was a netware server with dual Fijitsu 650Mb drives and a workstation with an AMD 486DX-133 overclocked. Although I'm not sure how much memory it had in it. At the time I had just upgraded a large number of workstations at work and had purchased 4.0 GB of ram (250x16Mb kits). All my home systems save one were bought and paid for by the company. I don't recall how much memory I had in my main workstation but it could have easily been 64Mb.
In any case, I think it was well within the realm of possibility for a non-average system to meet the hardware requirements to run a NNUE engine. Getting the software to do the same would have been the major problem I think.
Regards,
Zenmastur
Only 2 defining forces have ever offered to die for you.....Jesus Christ and the American Soldier. One died for your soul, the other for your freedom.
A version close to SF 11 was around 3035 at 20 kn/s.
A Pentium 75 would run SF 11 at speed around 6 kn/s. Rating for SF 11 on this P75 would be around 2900-2950.
SF 12 is around 50 RP over SF 11.
I'm afraid the network file may not fit on the harddisk of the i486.
No problem for the disk. Typical disk space at the time of the i486 was around 150 MB. At this time, I bought a 486-33 with a 120 MB disk and I added a 170 MB disk one or two year later.
But the problem would be the RAM, it was around 4 MB to 8 MB (space needed to run Windows 3.1 with some apps).
mclane wrote: ↑Tue Oct 13, 2020 12:06 pm
I would like to see Stockfish run on a 6502 so we can try out how it plays against the old 8 bit dedicated chess computers.
SF-NNUE depth 1 which is 500-1000 nodes is 2400Elo. That is at least 500 Elo above the best dedicated chess computers.
SF-NNUE on 6502 would have for sure more than 1000nodes per move. It wouldn't be any contest, it would be a slaughter.
Milos wrote: ↑Wed Oct 14, 2020 11:00 pm
SF-NNUE depth 1 which is 500-1000 nodes is 2400Elo. That is at least 500 Elo above the best dedicated chess computers.
SF-NNUE on 6502 would have for sure more than 1000nodes per move. It wouldn't be any contest, it would be a slaughter.
and how is that supposed to work? the 6502 has a 16-bit address space, even with MMU you'd have to map all the code + 20MB worth of weights.
plus with only 3 8-bit registers, well... sounds like a fairy tale
mar wrote: ↑Thu Oct 15, 2020 12:03 amand how is that supposed to work? the 6502 has a 16-bit address space, even with MMU you'd have to map all the code + 20MB worth of weights.
plus with only 3 8-bit registers, well... sounds like a fairy tale
I'm not sure this is what anyone was referring to, but given 2020's knowledge, it's plausible (likely, even) that simply shrinking the NNUE architecture down to suit whatever older hardware could handle would be the best thing you could do in any era.
Of course, most of SF-NNUE's heft is in the overparameterized input layer, so as you go back in time you will need to start being somewhat cleverer to further reduce a NNUE-like evaluation function's memory footprint, but nonetheless I suspect there shouldn't be any major impediments to getting something very like SF-NNUE running on 60's era computers.
Milos wrote: ↑Wed Oct 14, 2020 11:00 pm
SF-NNUE depth 1 which is 500-1000 nodes is 2400Elo. That is at least 500 Elo above the best dedicated chess computers.
SF-NNUE on 6502 would have for sure more than 1000nodes per move. It wouldn't be any contest, it would be a slaughter.
and how is that supposed to work? the 6502 has a 16-bit address space, even with MMU you'd have to map all the code + 20MB worth of weights.
plus with only 3 8-bit registers, well... sounds like a fairy tale
And SF-NNUE certainly won't reach 1000nps on a 6502 at 6502 speeds (1 or 2 mhz with maybe 1 instruction per 3 cycles, no multiplication and handling just 8 bits at a time).
It would be interesting to see what could be done with today's knowledge on, say, a C64, but the RAM limitation probably limits the applicability of many new techniques.
Last edited by syzygy on Thu Oct 15, 2020 7:55 pm, edited 1 time in total.
Milos wrote: ↑Wed Oct 14, 2020 11:00 pm
SF-NNUE depth 1 which is 500-1000 nodes is 2400Elo. That is at least 500 Elo above the best dedicated chess computers.
SF-NNUE on 6502 would have for sure more than 1000nodes per move. It wouldn't be any contest, it would be a slaughter.
and how is that supposed to work? the 6502 has a 16-bit address space, even with MMU you'd have to map all the code + 20MB worth of weights.
plus with only 3 8-bit registers, well... sounds like a fairy tale
And SF-NNUE certainly won't reach 1000nps on a 6502 at 6502 speeds (1 or 2 mhz with maybe 1 instruction per 3 cycles).
It would be interesting to see what could be done with today's knowledge on, say, a C64, but the RAM limitation probably limits the applicability of many new techniques.
Hehe, the SupraDrive hard-disk with 10 MB for 800,- Deutsch Mark appeared around 1986, not sure what the 20 MB version was, there you go with NNUE on 6502
smatovic wrote: ↑Thu Oct 15, 2020 7:53 pmHehe, the SupraDrive hard-disk with 10 MB for 800,- Deutsch Mark appeared around 1986, not sure what the 20 MB version was, there you go with NNUE on 6502
smatovic wrote: ↑Thu Oct 15, 2020 7:53 pmHehe, the SupraDrive hard-disk with 10 MB for 800,- Deutsch Mark appeared around 1986, not sure what the 20 MB version was, there you go with NNUE on 6502
Of course useless for NNUE if you still can't load the network into RAM.
Why not operate on the hard-disk, like Milos said, not nps but nodes per move, just a matter of time control, or alike. What where the tournament TCs back then?
