Evaluate river trail

[bhack]

I think that you can take advantages from the new Ecmascript proposal river trail instead of making plan on a not clear browser roadmap to support WebCL . Intel has already code on github at https://github.com/IntelLabs/RiverTrail/wiki

[bhack]

Take a look also at this benchmark https://github.com/Maratyszcza/blis-bench/blob/master/README.md and to this presentation http://www.cs.utexas.edu/users/flame/BLISRetreat2014/slides/marat-blis-retreat2014/

[bhack]

@karpathy what do you think?

[bhack]

Referencing node-forward/discussions#1

[tedmeeds]

Thanks, bhack for your suggestions. Yes, we would love to accelerate our code, refactor it, etc. If we find students to continue working on MLitB, acceleration will be a top priority. Ted.

[bhack]

Do you think that there could be some interaction between this and @karpathy convneths? In an interview @karpathy claimed: "Lastly, here's a crazy idea: massively distributed Neural Network training (think FoldIt, or SETI@Home), except every client merely visits a URL and right away starts to contribute Javascript compute time by sending gradient updates to a central server. A few issues have to be addressed first in terms of the modeling: vanilla Neural Networks have dense interactions so they are difficult to parallelize and naive use of distributed optimization techniques is likely to pose problems with stale gradients." My team is working in the caffe community and we have done some PR there. We are very interested in this kind of crowd computing large scale models. I don't know if a js solution that use parameter server in some way like non js peetum or a fully p2p can scale better with faster convergence.

[tedmeeds]

Sure, definitely. Can we take this discussion offline? Ted

[bhack]

Yes of course. If you want i can send you a mail..

[tedmeeds]

Thought you had it. [email protected]. Thanks.

[bjacob]

Don't hold your breath for widespread browser support for either WebCL (I don't know any browser vendor having plans to support it) or RiverTrail (I'm not aware of any browser vendor other than Mozilla showing an interest in it?)... instead, if you want something that runs in more than one browser, see how much you can do with JavaScript: Emscripten'ing C/C++ code into fast JavaScript, using Web Workers for multithreading, and if you need more power, look at SIMD.js and/or ongoing discussions to bring some pthreads-like multithreading capabilities to the Web.

[bhack]

There is an updated stage status at https://github.com/tc39/ecma262