You are viewing a plain text version of this content. The canonical link for it is here.

Posted to issues@systemml.apache.org by "Janardhan (JIRA)" <ji...@apache.org> on 2017/12/08 17:20:00 UTC

[jira] [Updated] (SYSTEMML-2041) Implement Block-Sparse GPU Kernels

     [ https://issues.apache.org/jira/browse/SYSTEMML-2041?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel ]

Janardhan updated SYSTEMML-2041:
--------------------------------
    Description: 
Sparsity enables, for example, training of neural networks that are much wider and deeper than otherwise possible with a given parameter budget and computational budget, such as LSTMs with tens of thousands of hidden units. (The largest LSTMs trained today are only thousands of hidden units.)

*Resource:* TensorFlow implemented repo - https://github.com/openai/blocksparse

*Best Supported architectures:* Maxwell, Pascal with Kepler & Volta support for limited functionality


  was:
Sparsity enables, for example, training of neural networks that are much wider and deeper than otherwise possible with a given parameter budget and computational budget, such as LSTMs with tens of thousands of hidden units. (The largest LSTMs trained today are only thousands of hidden units.)

*Resource:* TensorFlow implemented repo - https://github.com/openai/blocksparse



> Implement Block-Sparse GPU Kernels
> ----------------------------------
>
>                 Key: SYSTEMML-2041
>                 URL: https://issues.apache.org/jira/browse/SYSTEMML-2041
>             Project: SystemML
>          Issue Type: New Feature
>          Components: Infrastructure
>            Reporter: Janardhan
>         Attachments: GPU Kernels for Block-Sparse Weights.pdf
>
>
> Sparsity enables, for example, training of neural networks that are much wider and deeper than otherwise possible with a given parameter budget and computational budget, such as LSTMs with tens of thousands of hidden units. (The largest LSTMs trained today are only thousands of hidden units.)
> *Resource:* TensorFlow implemented repo - https://github.com/openai/blocksparse
> *Best Supported architectures:* Maxwell, Pascal with Kepler & Volta support for limited functionality



--
This message was sent by Atlassian JIRA
(v6.4.14#64029)