We've just released new improvements!
New AMD and Intel hardware targets are now available!
Intel Ice Lake - new!
Intel Cascade Lake - additional n2 instances
AMD EPYC Rome - new!
AMD EPYC Milan - additional n2 instances
INT8 Performance Improvement
OctoML regularly improves TVM performance on INT8 and FP32models. This week, we’ve released an update that improves AutoTVM performance on INT8 pre-quantized MobileNet by about 30% on x86 devices.
We have plans to continue improving performance on INT8 Resnet, Yolo, and SSD-MobileNet models on x86 and Arm Cortex A-72 devices in coming weeks.
TF SavedModels ( a broader superset of KerasSaved Models) are now available
In the Web UI, you will now see TF Saved Models a new option during model upload. Note: there will no longer be a separate option for Keras SavedModels, but you can upload Keras SavedModels under the TF SavedModels option.
This option will also be available in the SDK when you upgrade to the latest SDK version. Keras SavedModels will no longer be an option in the SDK starting 2.22.2022, so please upgrade your SDK!