Can't find it now... I've read a post about the Dolphin game emulator which started to use PEXT and BMI2. Performance was completely shot on AMD, and they now have multiple code paths.
edit: here...
https://dolphin-emu.org/blog/2020/02/07 ... -jan-2020/
Ryzen 2 and BMI2?
Moderators: hgm, Rebel, chrisw
-
- Posts: 12542
- Joined: Wed Mar 08, 2006 8:57 pm
- Location: Redmond, WA USA
Re: Ryzen 2 and BMI2?
I think more important is the number of benchmarks PDEP and PEXT take part in.
You don't want your numbers to look bad.
Taking ideas is not a vice, it is a virtue. We have another word for this. It is called learning.
But sharing ideas is an even greater virtue. We have another word for this. It is called teaching.
But sharing ideas is an even greater virtue. We have another word for this. It is called teaching.
-
- Posts: 710
- Joined: Sat Dec 06, 2014 1:53 pm
Re: Ryzen 2 and BMI2?
Ryzen 4700U + stockfish_20080822 abrok build
go to depth 28
Threads 1 vs 8
x64: 1 828 225 / 12 319 235
modern: 1 937 522 / 13 285 231
sse3: 1 821 351 / 12 494 956
bmi2: 1 256 031 / 9 255 682
avx2: 1 902 480 / 13 099 338
actual speed during long run will be lower, CPU mantain cTDP=25.0W (~3200 MHz @ 7 cores fishtest) for a little, then goes down to cTDP=15.0 (2300-2600 MHz) and stays there
go to depth 28
Threads 1 vs 8
x64: 1 828 225 / 12 319 235
modern: 1 937 522 / 13 285 231
sse3: 1 821 351 / 12 494 956
bmi2: 1 256 031 / 9 255 682
avx2: 1 902 480 / 13 099 338
actual speed during long run will be lower, CPU mantain cTDP=25.0W (~3200 MHz @ 7 cores fishtest) for a little, then goes down to cTDP=15.0 (2300-2600 MHz) and stays there
-
- Posts: 1564
- Joined: Thu Jul 16, 2009 10:47 am
- Location: Almere, The Netherlands
Re: Ryzen 2 and BMI2?
A few weeks ago I finally got my hand on a 10980XE for a reasonable price. It's speed is in the same ballpark as the AMD 3970X when I compare them with an equal number of cores. I prefer to use the 10980XE as my main workstation because it supports AVX-512 with VNNI, and the implementation of BMI2 (PEXT/PDEP) is much better. The 10980XE also supports Intel VTune which I used to use as my main profiler. I've degraded my 3970X to be a workhorse for bulk computations.mvanthoor wrote: ↑Tue Jun 02, 2020 3:23 pm I see, you're right. The 10940X is available now, for €899.
Now the question becomes: is it still worth it, to put such a CPU on a 3 year old chipset? If Intel doesn't keep the 2066 socket for the successor, The 10X-series is the end of the line for X299. Also, the CPU is quite expensive and "slow" compared to the AMD offerings, except for BMI2 and PEXT. The one thing I REALLY want is a 3950X with a good BMI2 and PEXT implementation.
The next ZEN iteration is planned for this fall. As I don't really NEED a new CPU right now, I should probably just wait and see if AMD fixes BMI2 and PEXT in that iteration, or if Intel introduces something new at the beginning of next year. Rumors of a new Z399 and X599 chipset for the X-cpu's have been circulating for over a year now.
I'm not one to wait for the next "big thing around the corner", except if it's so close (like the new ZEN iteration) that I'd hit my head against a wall repeatedly, if I bought a new computer now and the new ZEN's blow it out of the sky with a new BMI2 and PEXT implementation.
-
- Posts: 2661
- Joined: Wed Mar 10, 2010 10:18 pm
- Location: Hamburg, Germany
- Full name: Srdja Matovic
Re: Ryzen 2 and BMI2?
A NNUE benchmark on these AVX-512 2xFMA units would be interesting? I estimate it should operate then at least at 80% of SF speed.Joost Buijs wrote: ↑Sun Aug 09, 2020 7:21 pmA few weeks ago I finally got my hand on a 10980XE for a reasonable price. It's speed is in the same ballpark as the AMD 3970X when I compare them with an equal number of cores. I prefer to use the 10980XE as my main workstation because it supports AVX-512 with VNNI, and the implementation of BMI2 (PEXT/PDEP) is much better. The 10980XE also supports Intel VTune which I used to use as my main profiler. I've degraded my 3970X to be a workhorse for bulk computations.mvanthoor wrote: ↑Tue Jun 02, 2020 3:23 pm I see, you're right. The 10940X is available now, for €899.
Now the question becomes: is it still worth it, to put such a CPU on a 3 year old chipset? If Intel doesn't keep the 2066 socket for the successor, The 10X-series is the end of the line for X299. Also, the CPU is quite expensive and "slow" compared to the AMD offerings, except for BMI2 and PEXT. The one thing I REALLY want is a 3950X with a good BMI2 and PEXT implementation.
The next ZEN iteration is planned for this fall. As I don't really NEED a new CPU right now, I should probably just wait and see if AMD fixes BMI2 and PEXT in that iteration, or if Intel introduces something new at the beginning of next year. Rumors of a new Z399 and X599 chipset for the X-cpu's have been circulating for over a year now.
I'm not one to wait for the next "big thing around the corner", except if it's so close (like the new ZEN iteration) that I'd hit my head against a wall repeatedly, if I bought a new computer now and the new ZEN's blow it out of the sky with a new BMI2 and PEXT implementation.
--
Srdja
-
- Posts: 550
- Joined: Tue Nov 19, 2019 8:48 pm
- Full name: Alayan Feh
Re: Ryzen 2 and BMI2?
AVX-512 eats a lot of silicon die area, having a higher number of AVX2-capable cores instead would probably be better for many use cases.
-
- Posts: 1564
- Joined: Thu Jul 16, 2009 10:47 am
- Location: Almere, The Netherlands
Re: Ryzen 2 and BMI2?
Yes I think so, if you know how to do it. I didn't take a look at SF-NNUE yet, and I have no clue what it consists of.
-
- Posts: 1564
- Joined: Thu Jul 16, 2009 10:47 am
- Location: Almere, The Netherlands
Re: Ryzen 2 and BMI2?
Just downloaded the NNUE executables and one of the SV networks, each time I start an analysis the executable just quits without any message or warning. I tried several executables, so I guess I must be missing something.
I don't want to spend too much time on it because I have other things to do, maybe somebody has a suggestion?
I don't want to spend too much time on it because I have other things to do, maybe somebody has a suggestion?
-
- Posts: 2661
- Joined: Wed Mar 10, 2010 10:18 pm
- Location: Hamburg, Germany
- Full name: Srdja Matovic
Re: Ryzen 2 and BMI2?
I do not have the time either but if I would try to compile from source...Joost Buijs wrote: ↑Mon Aug 10, 2020 11:04 am Just downloaded the NNUE executables and one of the SV networks, each time I start an analysis the executable just quits without any message or warning. I tried several executables, so I guess I must be missing something.
I don't want to spend too much time on it because I have other things to do, maybe somebody has a suggestion?
I have read that some Xeons/Core-X have 2xFMA units, so I guess there must be
two of the AVX-512 units present. If right this means that NNUE does not only
profit from 512 bit vector instructions but also from SMT/HyperThreading able
to utilize two vector units. NNUE running on such a Xeon machine could surpass
SF speed, maybe we will see such a thing happening on TCEC/CCC, not sure what
hardware they currently use. Anyway, maybe I will take a look what hardware
Google/Amazon Cloud currently offers and make a benchmark some day...
--
Srdja
-
- Posts: 1784
- Joined: Wed Jul 03, 2019 4:42 pm
- Location: Netherlands
- Full name: Marcel Vanthoor
Re: Ryzen 2 and BMI2?
Cool I'm in doubt if I should spend any money now. Due to my vision problem (caused by cataract, which is due for surgery at the end of the month) development of my chess engine (and my attempt to restart PicoChess development) have virtually come to a halt for now. Apart from that, I don't want to build an Intel 10x-based computer onto the X299 chipset, and for the development I'm doing now, the old 6700K is still good enough.Joost Buijs wrote: ↑Sun Aug 09, 2020 7:21 pm A few weeks ago I finally got my hand on a 10980XE for a reasonable price. It's speed is in the same ballpark as the AMD 3970X when I compare them with an equal number of cores. I prefer to use the 10980XE as my main workstation because it supports AVX-512 with VNNI, and the implementation of BMI2 (PEXT/PDEP) is much better. The 10980XE also supports Intel VTune which I used to use as my main profiler. I've degraded my 3970X to be a workhorse for bulk computations.
I'll wait for something from Intel in 2021. I refuse to buy a CPU with crippled BMI2/PEXT execution if that is the one thing I actually NEED for my personal programming project