[Embench] Candidate for ML for future Embench

Mark Hill mark.hill at huawei.com
Wed Apr 22 16:34:39 CEST 2020


I think opus encode/decode could also be  interesting benchmarks to include, especially as opus can be compiled to either use floating point or fixed point calculations (using purely integer arithmetic). I compiled the opus_demo.c and it is around 300K of code, converting this to a standalone form suitable for inclusion in EMBENCH is straightforward, it just needs the file-to-file conversion to be converted to read/write a couple of data stuctures instead.

https://github.com/xiph/opus

Mark


From: Embench [mailto:embench-bounces at lists.librecores.org] On Behalf Of David PATTERSON
Sent: 21 April 2020 17:30
To: embench at lists.librecores.org
Subject: [Embench] Candidate for ML for future Embench

MLPerf's TinyML benchmark effort is considering using this model for visual wake words. The NN model is 250 KB and they are use TensorFlow Lite.
https://github.com/tensorflow/tensorflow/tree/master/tensorflow/lite/micro/examples/person_detection

Is this size plausible for future benchmarks?

Dave
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.librecores.org/pipermail/embench/attachments/20200422/c7fe6f7d/attachment.htm>


More information about the Embench mailing list