Is Blas, or OpenCL really the right way to go

[AdminX]

I've been wondering is Blas or OpenCL really the right way to go for LC0 on CPU? Because when I look at the performance of Komodo MCTS on CPU, Komodo appears to be light years ahead of LC0 as far as MCTS is concerned. Of course Komodo is no match for LC0 on GPU, but on CPU, LC0 would get smashed by Komodo.

[dkappe]

I've been wondering is Blas or OpenCL really the right way to go for LC0 on CPU? Because when I look at the performance of Komodo MCTS on CPU, Komodo appears to be light years ahead of LC0 as far as MCTS is concerned. Of course Komodo is no match for LC0 on GPU, but on CPU, LC0 would get smashed by Komodo.

You are assuming that Komodo MCTS is using a NN. I don’t believe that is the case.

[AdminX]

I've been wondering is Blas or OpenCL really the right way to go for LC0 on CPU? Because when I look at the performance of Komodo MCTS on CPU, Komodo appears to be light years ahead of LC0 as far as MCTS is concerned. Of course Komodo is no match for LC0 on GPU, but on CPU, LC0 would get smashed by Komodo.

You are assuming that Komodo MCTS is using a NN. I don’t believe that is the case.

No I am not. But I am wondering if Blas or OpenCL are really the best options for CPU that LC0 has.

[Robert Pope]

I've been wondering is Blas or OpenCL really the right way to go for LC0 on CPU? Because when I look at the performance of Komodo MCTS on CPU, Komodo appears to be light years ahead of LC0 as far as MCTS is concerned. Of course Komodo is no match for LC0 on GPU, but on CPU, LC0 would get smashed by Komodo.

You are assuming that Komodo MCTS is using a NN. I don’t believe that is the case.

No I am not. But I am wondering if Blas or OpenCL are really the best options for CPU that LC0 has.

Then what does comparing it to Komodo MCTS have to do with anything?

[dkappe]

You are assuming that Komodo MCTS is using a NN. I don’t believe that is the case.

No I am not. But I am wondering if Blas or OpenCL are really the best options for CPU that LC0 has.

OK. You had me confused.

If we use a NN for leela, then the best way of computing that is a really, really fast linear algebra library. On CPU, blas is one of the top performers.

Now if you want to do something other than blas, then you’ll have to abandon the NN. Maybe use something like a shallow ab search to drive the mcts. But at this point it’s a very different animal than leela.

[AdminX]

You are assuming that Komodo MCTS is using a NN. I don’t believe that is the case.

No I am not. But I am wondering if Blas or OpenCL are really the best options for CPU that LC0 has.

OK. You had me confused.

If we use a NN for leela, then the best way of computing that is a really, really fast linear algebra library. On CPU, blas is one of the top performers.

Now if you want to do something other than blas, then you’ll have to abandon the NN. Maybe use something like a shallow ab search to drive the mcts. But at this point it’s a very different animal than leela.

Thank You, So for people with Haswell or later Intel CPU's does MKL perform better much better than than the Blas only version or just slightly so? I have not been testing the CPU versions all that much untill resently, but would enjoy getting the most out of the CPU version as I can.

[Dann Corbit]

It is an important and interesting question.
Here is a benchmark link for various blas type bundles:
https://medium.com/datathings/benchmark ... 57fb1c6dc7

[Dann Corbit]

Also interesting is the cost per dollar calculation for various cards here:
https://timdettmers.com/2019/04/03/whic ... -learning/

[Max]

Recently compiled Lc0 0.21.2 with the three different blas variants on my mac and run a small test on network 11258-32x4-se.pb.gz.

My MacBook Air with Intel Core i5 from 2015 shows after "go nodes 50000" about

• 2400 nps with Apple vecLib, thread=1
  3500 nps with Apple vecLib, threads=2
  
  2350 nps with openBLAS 0.3.6 from Homebrew, thread=1
  4000 nps with openBLAS 0.3.6 from Homebrew, threads=2
  
  2750 nps with MKL 2019u4 from Intel, thread=1
  4600 nps with MKL 2019u4 from Intel, threads=2

With network netT40.T8.610 only ~50 nps are possible with BLAS.

On my MacBook Air the Intel HD6000 gpu acts slower with openCL. Not to mention the 35+ watts are "killing" the notebook within seconds. Using Intel Power Gadget for this scenario seems an absolute must.

[AdminX]

Thank Dann,

I was aware of the second link as I saw it posted on Discord, but not the first. Good info.