Announcements

OpenMP will be disabled in future official builds (build option will still be available). A NoOpenMP version of ONNX Runtime is now available with this release on Nuget and PyPi for C/C++/C#/Python users.
In the next release, MKL-ML, openblas, and jemallac build options will be removed, and the Microsoft.ML.OnnxRuntime.MKLML Nuget package will no longer be published. Users of MKL-ML are recommended to use the Intel EPs. If you are using these options and identify issues switching to an alternative build, please file an issue with details.

Key Feature Updates

ONNX 1.8 support / opset 13
New contrib ops: BiasSoftmax, MatMulIntegerToFloat, QLinearSigmoid, Trilu
ORT Mobile now compatible with NNAPI for accelerating model execution on Android devices
Build support for Mac with Apple Silicon (CPU only)
New dependency: flatbuffers
Support for loading sparse tensor initializers in pruned models
Support for setting the execution priority of a node
Support for selection of cuDNN conv algorithms
BERT Model profiling tool

New session option to disable denormal floating numbers on sse3 supporting CPUs
- Eliminates unexpected performance degradation due to denormals without needing to retrain the model
Option to share initializers between sessions to improve memory utilization
- Useful when several models that use the same set of initializers except the last few layers of the model are loaded in the same process
- Eliminates wasteful memory usage when every model (session) creates a separate instance of the same initializer
- Exposed by the AddInitializer API
Transformer model optimizations
- Longformer: LongformerAttention CUDA operator added
- Support for BERT models exported from Tensorflow with 1 or 2 inputs
- Python optimizer supports additional models: openai-GPT, ALBERT and FlauBERT
Quantization
- Support of per-channel QuantizeLinear and DeQuantizeLinear
- Support of LSTM quantization
- Quantization performance improvement on ARM
- CNN quantization perf optimizations, including u8s8 support and NHWC transformer in QLinearConv
ThreadPool
- Use _mm_pause() for spin loop to improve performance and power consumption

DNNL EP Updates
- DNNL updated from 1.1.1 to 1.7
NNAPI EP Updates
- Support for CNN models
- Additional operator support - Resize/Flatten/Clip
TensorRT EP Updates
- Int8 quantization support (experimental)
- Engine cache refactoring and improvements
- General fixes and performance improvements
OpenVINO EP Updates
- OpenVINO 2021.1 support
- OpenVINO EP builds as shared library
- Multi-threaded inferencing support
- fp16 input type support
- Multi-device plugin support
- Hetero plugin support
- Enable build on ARM64
DirectML EP Updates (1.3.0 -> 1.4.0)
- Utilizing the first public standalone release of the DirectML API through the DirectML NuGet package release
- General fixes and improvements
nGraph EP is removed. Recommend to use OpenVINO instead

VCRuntime2019 with OpenMP: pinning a process to NUMA node 1 forces the execution to be single threaded. Fix is in progress in VC++.
- Workaround: place the VS2017 vcomp DLL side-by-side so that ORT uses the VS2017 version
Pip version >=20.3 is required for use on macOS Big Sur (11.x)
The destructor of OrtEnv is now non-trivial and may do DLL unloading Do not call ReleaseEnv from DLLMain or put OrtEnv in global variables. It is not safe to call FreeLibrary from DllMain. - reference
Some unit tests fail on Pascal GPUs. See: #5914
If using the default CPU package (built with OpenMP), consider tuning the OpenMP settings to improve performance. By default the number of threads to use for openmp parallel regions is set to the number of logical CPUs. This may not be optimal for machines with hyper-threading; when CPUs are oversubscribed the 99-percentile latency could be 10x greater. Setting the OMP_NUM_THREADS environment variable to the number of physical cores is a good starting point. As noted in Announcements, future official builds of ORT will be published without OpenMP