The loss of speed is expected due to the larger network. I am not sure what exactly is going on and your results seem to match (the elo difference to 6.23 and 6.0). I was just curious why we are struggling to beat SF 10 whereas it beats SF 11 on my machine.
Just 200 games statistically is not very sound. I started a match, Koivisto 7 vs SF10, normal openings, 1000 games.
The loss of speed is expected due to the larger network. I am not sure what exactly is going on and your results seem to match (the elo difference to 6.23 and 6.0). I was just curious why we are struggling to beat SF 10 whereas it beats SF 11 on my machine.
Just 200 games statistically is not very sound. I started a match, Koivisto 7 vs SF10, normal openings, 1000 games.
AndrewGrant wrote: ↑Tue Jul 06, 2021 10:30 pmI would love to see a push for all rating lists to make use of AVX2 processors. Most NNUE engines are gimped without those.
I fear another PC in the living room will be at the cost my marriage.
My charm won
BTW, hardly any speed difference between ethereal 13.25 NNUE AVX2 and the popcount version.
90% of coding is debugging, the other 10% is writing bugs.
AndrewGrant wrote: ↑Tue Jul 06, 2021 10:30 pmI would love to see a push for all rating lists to make use of AVX2 processors. Most NNUE engines are gimped without those.
I fear another PC in the living room will be at the cost my marriage.
My charm won
BTW, hardly any speed difference between ethereal 13.25 NNUE AVX2 and the popcount version.
Maybe not on your hardware because you don't have avx2 instructions on your hardware, but other people with avx2 instructions on their hardware can detect and have detected a speed difference.
AndrewGrant wrote: ↑Tue Jul 06, 2021 10:30 pmI would love to see a push for all rating lists to make use of AVX2 processors. Most NNUE engines are gimped without those.
I fear another PC in the living room will be at the cost my marriage.
My charm won
BTW, hardly any speed difference between ethereal 13.25 NNUE AVX2 and the popcount version.
Maybe not on your hardware because you don't have avx2 instructions on your hardware, but other people with avx2 instructions on their hardware can detect and have detected a speed difference.
You are wrong, an AVX2 compile will crash if the processor doesn't support AVX2.
90% of coding is debugging, the other 10% is writing bugs.
AndrewGrant wrote: ↑Tue Jul 06, 2021 10:30 pmI would love to see a push for all rating lists to make use of AVX2 processors. Most NNUE engines are gimped without those.
I fear another PC in the living room will be at the cost my marriage.
My charm won
BTW, hardly any speed difference between ethereal 13.25 NNUE AVX2 and the popcount version.
Maybe not on your hardware because you don't have avx2 instructions on your hardware, but other people with avx2 instructions on their hardware can detect and have detected a speed difference.
You are wrong, an AVX2 compile will crash if the processor doesn't support AVX2.
Are you using an AMD processor? The problem is that while older Ryzen processors "support" AVX2, it's not a hardware instruction. It's done in software by dividing it in 2 and using normal AVX to process it, so, while it's supported, it works exactly the same (speed wise) as normal AVX so there's no boost in speed. In Intel processors (and Ryzen 3000 and newer), avx2 is real and you see almost a 100% speed increase.
CMCanavessi wrote: ↑Wed Nov 03, 2021 7:16 pm
Are you using an AMD processor? The problem is that while older Ryzen processors "support" AVX2, it's not a hardware instruction. It's done in software by dividing it in 2 and using normal AVX to process it, so, while it's supported, it works exactly the same (speed wise) as normal AVX so there's no boost in speed. In Intel processors (and Ryzen 3000 and newer), avx2 is real and you see almost a 100% speed increase.
Sorry to ask, and how do you know about this? I have an AMD processor - how to check if it is fully used AVX2.
AndrewGrant wrote: ↑Tue Jul 06, 2021 10:30 pmI would love to see a push for all rating lists to make use of AVX2 processors. Most NNUE engines are gimped without those.
I fear another PC in the living room will be at the cost my marriage.
My charm won
BTW, hardly any speed difference between ethereal 13.25 NNUE AVX2 and the popcount version.
Maybe not on your hardware because you don't have avx2 instructions on your hardware, but other people with avx2 instructions on their hardware can detect and have detected a speed difference.
You are wrong, an AVX2 compile will crash if the processor doesn't support AVX2.
Are you using an AMD processor? The problem is that while older Ryzen processors "support" AVX2, it's not a hardware instruction. It's done in software by dividing it in 2 and using normal AVX to process it, so, while it's supported, it works exactly the same (speed wise) as normal AVX so there's no boost in speed. In Intel processors (and Ryzen 3000 and newer), avx2 is real and you see almost a 100% speed increase.
Intel.
90% of coding is debugging, the other 10% is writing bugs.
CMCanavessi wrote: ↑Wed Nov 03, 2021 7:16 pm
Are you using an AMD processor? The problem is that while older Ryzen processors "support" AVX2, it's not a hardware instruction. It's done in software by dividing it in 2 and using normal AVX to process it, so, while it's supported, it works exactly the same (speed wise) as normal AVX so there's no boost in speed. In Intel processors (and Ryzen 3000 and newer), avx2 is real and you see almost a 100% speed increase.
Sorry to ask, and how do you know about this? I have an AMD processor - how to check if it is fully used AVX2.
Easy, if you have a Zen 2 processor or newer, you're good to go. Zen 2 processors are Ryzen 3xxx and newer. If you have a Zen 1 or Zen + like I do (Ryzen 1xxx or 2xxx) you're busted