What’s New¶
Main release notes¶
This release (Neuron 1.13.0) introduces higher performance , updated frameworks support, additional tutorials, and adding models and tools:
Additional compiler improvements boost performance up to 20% higher throughput compared to previous release across model types.
Improving usability for NLP models, with out-of-the-box 12x higher-throughput at 70% lower cost for Hugging Face Transformers pre-trained BERT Base models, see Using NeuronCore Pipeline with PyTorch Tutorial.
Upgrade Apache MXNet (Incubating) to 1.8, where Neuron is now a plugin, see Neuron Apache MXNet (Incubating) Release Notes.
PyTorch ResNext models now functional with new operator support, see PyTorch Neuron release notes.
PyTorch Yolov5 support, see PyTorch Neuron release notes.
MXNet (Incubating): Gluon API and Neuron support for NLP BERT models, see Neuron Apache MXNet (Incubating) Release Notes.
PyTorch Convolution operator support has been extended to include most Conv1d and Conv3d variants, please see Supported operators [PyTorch] for the complete list of operators.
First release of Neuron plugin for TensorBoard, see Neuron Plugin for TensorBoard Release Notes.
Software Deprecation
Detailed release notes¶
Software |
Details |
|---|---|
Main |
|
PyTorch |
|
TensorFlow |
|
Apache MXNet (Incubating) |
|
Compiler |
|
Runtime |
|
Tools |
|
DLAMI and Conda Packages |
|
Software Deprecation |