2018-08-19

Lc0 v0.17.0-rc1 has been released.

The release candidate of a new version of the Lc0 engine has been released.

v0.17.0-rc1

We expect to have a stable v0.17.0 release in one week, so that we can use it for CCCC. For now you can either help us to find bugs by trying the RC1, or use v0.16 for now.

Download and full changelog here.


Change highlights:

  • Syzygy Tablebases support.
    Only WDL probe for now, e.g. the engine only probes for positions after pawn moves and captures.
  • Ponder support.
  • Batch support for OpenCL backend, that gives up to 5x speedup.

    UPD:It turned out that openCl batching is off by default, so no 5x speedup is visible.
    To enable, use command like flag --backend_opts=batch_size=16.
    Due to another issue, it may happen that batch size 16 requires too much VRAM so it doesn't start. In that case, try lower values.
  • Windows CUDA version of Lc0 now includes all required .dlls.

14 comments:

  1. I have question about the Batch support for OpenCL backend. Is it exclusive for Nvidia Cards? Is the 5x speedup means compared to CUDA version?

    ReplyDelete
    Replies
    1. No, it also works on other GPUs with OpenCl support (i.e. AMD). 5x is speedup compared to OpenCl in v0.16.
      It's still significantly slower than CUDA version.

      Delete
    2. If you have a Nvidia card, then you should not use OpenCL (unless your Nvidia card is so old that it does not support CUDNN 7.1). So, the OpenCL speedup is not relevant if you have a Nvidia card.

      Delete
  2. So is CUDA or OpenCL version being used at CCCC?

    ReplyDelete
  3. Hash problem solved? when i use for example in fritz gui hash 2GB in few moves it shows hash full 100% and do not reset to 0% after each move as other engines. Also until lc0 start thinking there is about 5 second, other engines start thinking immediately.

    ReplyDelete
    Replies
    1. Hash usage display (which is actually cache usage display) problem is not actually a problem. Cache works properly and older entries are evicted as soon as there is something fresh to insert. It doesn't cause any slowdown or loss of strength, only causes confusion to whoever looks at it.

      Delete
  4. are there binaries for windows available?

    ReplyDelete
    Replies
    1. Yes, here: https://github.com/LeelaChessZero/lc0/releases/tag/v0.17.0-rc1

      Delete
  5. Hi, just a quick question; Are tensor cores utilized? 4xV100 with tensor cores would be sick on CCCC!

    ReplyDelete
    Replies
    1. Currently lc0 can only gather batch of 300-500 at most, and it sends minibatch of size 256 to GPU. That means, that it can hardly utilize 2 GPUs and more than that is kind of useless.
      We have ideas how to improve that, not entirely sure whether they will work though.

      Delete
  6. Fuck yeah boys our compy time is paying off!

    ReplyDelete
  7. tablebases dont work in fritz gui

    ReplyDelete
  8. why lc0 at cccc use poor 11089 network and not one of best one like 11189 ?

    ReplyDelete