mxnet-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Naveen Swamy <mnnav...@gmail.com>
Subject Re: Include MKLDNN into default mxnet pip package
Date Thu, 22 Nov 2018 01:08:14 GMT
Hi Alex,

Thanks for promptly running the numbers on AMD and reporting here.

Can you please update the AMD numbers here for posterity
https://cwiki.apache.org/confluence/display/MXNET/MXNet+with+Intel+MKL-DNN+-+Performance+Benchmarking
?

are there any outstanding issues when MKLDNN is enabled? from my offline
conversation I am briefly aware performance issues with LSTM, is there an
GitHub issue for it?

MKLDNN is a submodule dependency, are we pulling the latest commit or
releases  ? If not we should move to releases before we make it a default.
Ideally we should use platform specific distributions (-dev packages) at
least we should rely on well tested releases.


Thanks, Naveen

On Wed, Nov 21, 2018 at 4:55 PM Zai, Alexander <alexzai@amazon.com.invalid>
wrote:

> AMD benchmarks have been published. We are seeing a x15.8 speedup with
> Resnet50 (batch size 32) on AWS's new m5a.24xlarge machine. With a smaller
> network (Mobilenet - batch size 32) the speedup is more significant at
> x38.7. Let's have a vote to see if the PR to have MKLDNN enabled by default
> (https://github.com/apache/incubator-mxnet/pull/12591) can be merged
> before 1.4.0 release.
>
> On 10/19/18, 9:17 AM, "Pedro Larroy" <pedro.larroy.lists@gmail.com>
> wrote:
>
>     I did  pip install mxnet-mkl==1.3.1b20181018 on an AMD Ryzen 1950X and
> unit
>     tests are passing.
>
>     Is this build using AVX512?  in /proc/cpuinfo I see only "avx" flag.
>     There's no "avx2" like on recent intel cpus.
>
>     Pedro.
>
>     On Fri, Oct 19, 2018 at 5:12 PM Hagay Lupesko <lupesko@gmail.com>
> wrote:
>
>     > Awesome collaborative effort across many contributors and companies!
>     >
>     > The boost is impressive and for MXNet users to get this boost "out
> of the
>     > box" is a great benefit and makes MXNet an even better choice.
>     >
>     > Alex - can you clarify whether there are any down sides with regards
> to
>     > noon AVX-512 architectures, AMD CPUs, etc? Will it gracefully
> fallback?
>     >
>     > Hagay
>     >
>     >
>     > On Fri, Oct 19, 2018, 15:46 Sergio Fernández <wikier@apache.org>
> wrote:
>     >
>     > > If there is no downside on platforms not supporting AVX512
> instructions,
>     > > then +1
>     > >
>     > >
>     > > On Wed, Oct 17, 2018, 14:10 Alex Zai <azai91@gmail.com> wrote:
>     > >
>     > > > Hey all,
>     > > > We have been working hard these past few months to integrate and
>     > > stabilize
>     > > > Intel’s MKLDNN deep learning CPU accelerator into Mxnet and have
> made
>     > > > incredible progress. On CPUs with AVX512 instructions (such as
> c5.18x)
>     > we
>     > > > have seen performance increase up to 12x and on other platforms
> (Macs,
>     > > > AVX2) we seen a speedup of 1.5+. Full list of benchmarks can be
> found
>     > > here
>     > > > (
>     > > >
>     > >
>     >
> https://cwiki.apache.org/confluence/pages/viewpage.action?pageId=95650764
>     > > >  and https://github.com/apache/incubator-mxnet/pull/12591).
>     > > >
>     > > > Currently, using this accelerator requires the developer to
> either pip
>     > > > install the mxnet-mkl version of mxnet or to build it themselves
> from
>     > > > source. Given that we should try to provide the best performance
> "out
>     > of
>     > > > the box” with mxnet we should include this in the default build.
> The
>     > > mkldnn
>     > > > library is included with in the pip package build so it does not
>     > require
>     > > an
>     > > > external dependency.
>     > > >
>     > > > There were concerns that MKLDNN could cause regressions on
> certain
>     > > > platforms (as it did with the tensorflow version a while back);
> but we
>     > > > added a env flag (MXNET_MKLDNN_ENABLED) that allows users to
> turn of
>     > this
>     > > > feature during runtime. Please bring up any other concerns you
> may have
>     > > and
>     > > > your thoughts on including this accelerator in the default build.
>     > > >
>     > > > Best,
>     > > > Alex
>     > > >
>     > >
>     >
>
>
>

Mime
  • Unnamed multipart/alternative (inline, None, 0 bytes)
View raw message