Overview
This document provides details about new features and known issues for the Intel® AI Tools. The tools includes the following components:
-
Intel® Extension for TensorFlow*
-
Intel® Extension for PyTorch*
-
Intel® Distribution for Python*
-
Intel® AI Reference Models
-
Intel® Distribution of Modin*
Where to Find the Release
Please see the AI Tools Selector Guide for more information on how to acquire the package.
Compatibility Notes
Component | Compatibility Version |
Intel® Extension for TensorFlow* | 2.15 |
PyTorch* Optimizations from Intel | 2.3 (CPU), 2.1.40 (XPU) |
Intel® Distribution for Python* | 3.9 |
Intel® Distribution of Modin* | 0.30.0 |
Intel® Neural Compressor | 2.5.1 |
Intel® AI Reference Models | 3.2 |
What's New in Intel® AI Tools 2024.2
- oneDNN provides enhanced Performance for next generation client platforms: Experience faster and more efficient processing with broad production quality optimizations, maximizing the performance potential of next gen AI enhanced client platforms.
- oneDNN provides optimized Performance for next generation server platforms: Future-proof your systems with enhanced production quality optimizations, ensuring top-tier performance for next gen Intel Xeon Scalable processors.
- In the latest update, the Intel® Distribution of Modin has been integrated with the conda-forge distribution. Moving forward, Intel will contribute optimizations directly to the upstream distribution and will no longer maintain a separate Modin package on the conda Intel channel. For optimal performance, we advise using Modin from conda-forge, complemented by performance-enhancing libraries and compiler runtimes available from the Intel channel.
- To extend sparsity functions across oneDAL algorithms, this release adds DPC++ sparse gemm and gemv primitives and includes sparsity support for the logloss function primitive.
- Intel® Extension for Scikit-learn introduces the IncrementalLinearRegression interface so that each algorithm iteration is based on the previously changed output data, reducing the computational costs for users by making computations only on changed dataset pieces rather than the full one.
- Intel® Neural Compressor delivers new pruning and optimization in this release.
System Requirements
Please see system requirements.
How to Start Using the Tools
Please reference the usage guides for each of the included tools:
- Intel® Optimization for TensorFlow* - See Performance Considerations
- Intel® Optimization for PyTorch * - See Performance Considerations
Compatibility Notes
Component | Compatibility Version |
Intel® Extension for TensorFlow* | 2.15 |
PyTorch* Optimizations from Intel | 2.2 (CPU), 2.1.20 (XPU) |
Intel® Distribution for Python* | 3.9 |
Intel® Distribution of Modin* | 0.26.1 |
Intel® Neural Compressor | 2.4.1 |
Intel® AI Reference Models | 3.1 |
What's New in Intel® AI Tools 2024.1
- Unlock performance enhancements with the latest Intel® oneAPI Deep Neural Network Library (oneDNN) including improvements in graphics processing for Intel data center GPUs and Intel® Arc™ Graphics, perfect for complex models like Large Language Models and Stable Diffusion and increased performance for Intel Xeon Scalable processors.
- Introducing the latest 2024.1 release of the Intel Distribution of Modin, delivering significant enhancements in both security and performance. A robust security solution ensures proactive identification and remediation of vulnerabilities, offering protection for your organization's data assets. Additionally, we've implemented numerous performance fixes to optimize asynchronous execution, resulting in improved efficiency across operations.
- Speed up gradient boosting inference across XGBoost, LightGBM, and CatBoost* without sacrificing accuracy with new fast tree inference in Intel® oneAPI Data Analytics Library (oneDAL).
- Perform parallel computations at reduced cost on Intel GPUs with new Intel® Extension for Scikit-Learn* algorithms.
-
Get your AI Tools faster with the new web-based AI Tools Selector, now including OpenVINO™ toolkit
Compatibility Notes
Component | Compatibility Version |
Intel® Extension for TensorFlow* | 2.14 |
PyTorch* Optimizations from Intel | 2.0.12 |
Intel® Distribution for Python* | 3.9 |
Intel® Distribution of Modin* | 0.24 |
Intel® Neural Compressor | 2.3 |
Intel® AI Reference Models | 3.0 |
What's New in Intel® AI Tools 2024.0.2
-
"Intel® Analytics Toolkit” renamed to “Intel® AI Tools"
-
AI Tools Selector allows you to select the tools you need, instead of the whole tools.
-
You can still download the offline installer to grab all the tools.
-
Supported package types Conda, pip and docker, with the others like apt and yum not supported.
-
Intel® Distribution for Python* is now a separate download and not offered in the Selector.
-
Model Zoo changes its name and it is not available in the Selector.
-
Selector only supports Linux.
-
oneCCL bug fixes.
Compatibility Notes
Component | Compatibility Version |
Intel® Optimization for TensorFlow* | 2.13 |
Intel® Extension for TensorFlow* | 2.13 |
Intel® Optimization for PyTorch* | 2.1 |
Intel® Extension for PyTorch* | 2.0.110 |
Intel® oneCCL Bindings for PyTorch* | 2.0.110 |
Intel® Distribution for Python* | 3.9 |
Intel® Distribution of Modin* (Third Party Program) | 0.23.0 |
Intel® Neural Compressor | 2.1 |
Model Zoo for Intel® Architecture | 2.12 |
What's New in AI Analytics Toolkit 2023.2
-
The NEW Intel® AI Tools Selector beta to discover and install Intel® AI tools, libraries, and frameworks with greater flexibility to meet your needs, including recommendations to match common workflows and custom configurations.
-
Enhanced TensorFlow performance through a reduction in the number of “cast” operations for bfloat16 models using Intel® Advanced Matrix Extensions (Intel® AMX) with the Intel® Extension for TensorFlow.
-
Accelerate data preprocessing with Pandas 2.0 support in the Intel® Distribution of Modin, combining the latest advancements in Pandas with the benefits of parallel and distributed computing.
-
Faster training and inference for AI workloads, increased hardware choice, simpler debug and diagnostics, and support for graph neural network processing, Intel® oneAPI Deep Neural Network Library (oneDNN) now has improved performance optimization for 4th Gen Intel® Xeon® Processors and Intell® GPUs, extended diagnostics in verbose mode, and experimental support of Graph Compiler backend for Graph API.
.
System Requirements
Please see system requirements.
Please reference the usage guides for each of the included tools:
- Intel® Optimization for TensorFlow* - See Performance Considerations
- Intel® Optimization for PyTorch * - See Performance Considerations
Known Limitations
- Utilizing Intel® Optimization for PyTorch and Intel® Extension for PyTorch on Intel® Xeon® Scalable Processors Familiy without GPU accelerator requires installing Intel® Extension for PyTorch with CPU support inside pytorch conda environment with following instructions
source /opt/intel/oneapi/setvars.sh
conda create --name pytorch-cpu --clone pytorch
conda activate pytorch-cpu
conda remove intel-extension-for-pytorch pytorch oneccl_bind_pt torchvision --force -y
python -m pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cpu
python -m pip install intel-extension-for-pytorch oneccl_bind_pt --extra-index-url https://pytorch-extension.intel.com/release-whl/stable/cpu/us/
- No attribute 'convert_variables_to_constants' from Intel® Neural Compressor
- In cases that an error "AttributeError: module 'tensorflow.python.framework.graph_util' has no attribute 'convert_variables_to_constants' is showned during TensorFlow Model Quantization via INC, please use PB format model instead of model checkpoint
- No module named 'keras.layers.convolutional' from Intel® Neural Compressor
- There is incompatibility between INC 2.2 and keras 2.13. As a workaround user might install keras=2.12 and keras quantization will pass.
- Potential installation failure
- Use of the "download only" option at install time may result in an "Installation error" due to "Integrity check failed". The cause is an uncommon race condition in the installation process. Typically, executing the installation again will succeed.
Compatibility Notes
Component | Compatibility Version |
Intel® Optimization for TensorFlow* | 2.12 |
Intel® Extension for TensorFlow* | 1.2 |
Intel® Optimization for PyTorch* | 1.13 |
Intel® Extension for PyTorch* | 1.13.120 |
Intel® oneCCL Bindings for PyTorch* | 1.13.120 |
Intel® Distribution for Python* | 3.9 |
Intel® Distribution of Modin* | 0.19.0 |
Intel® Neural Compressor | 2.1 |
Model Zoo for Intel® Architecture | 2.11 |
What's New in AI Analytics Toolkit 2023.1.1
- Intel® Neural Compressor optimizes auto-/multinode-tuning strategy & LLM memory.
- Intel® Distribution of Modin introduces a new, experimental NumPy API.
- Model Zoo for Intel® Architecture adds support for dataset downloader and data connectors.
- Intel® Extension for TensorFlow now supports TensorFlow 2.12 and adds Ubuntu 22.04 and Red Hat Enterprise Linux 8.6 to the list of supported platforms.
- Intel® Extension for PyTorch is now compatible with oneDNN 3.1 which improves on PyTorch 1.13 operator coverage..
System Requirements
Please see system requirements.
Please reference the usage guides for each of the included tools:
- Intel® Optimization for TensorFlow* - See Performance Considerations
- Intel® Optimization for PyTorch * - See Performance Considerations
Known Limitations
- Runtime Out of Memory Error Using GPU
-
There is a potential known issue with the cleanup of resources at queue synchronization points in longer running jobs (most likely to show up in multi-tile or multi-device setups) that can lead to resources on the device being used up and causing out of memory errors.
-
Workaround
-
In cases where this is identified, users can use the compiler hotfix available at below links to help address this situation. With this hotfix in place, cleanup of these resources will happen more often, i.e. after certain threshold of the number of allocated resources is hit. While a default value for this has been provided, users can enforce finer grain control through the usage of the environmental variable SYCL_PI_LEVEL_ZERO_COMMANDLISTS_CLEANUP_THRESHOLD. The value is defined as: If non-negative, then the threshold is set to this value. If negative, the threshold is set to INT_MAX. Whenever the number of command lists in a queue exceeds this threshold, an attempt is made to cleanup completed command lists for their subsequent reuse. The default is 20
-
Compiler hotfix
-
-
Compatibility Notes
-
Intel® Optimization for TensorFlow* is compatible with version 2.10.0
-
Intel® Extension for TensorFlow* is compatible with version 1.1.0
-
Intel® Optimization for PyTorch* is compatible with version 1.13.0
-
Intel® Extension for PyTorch* is compatible with version 1.13.10
-
Intel® oneCCL Bindings for PyTorch* is compatible with version 1.13.0 (cpu) or 1.13.100 (gpu)
-
Intel® Distribution for Python* is compatible with cpython version 3.9
-
Intel® Distribution of Modin* is compatible with version 0.17
-
Intel® Neural Compressor is compatible with version 1.14.2
-
Model Zoo for Intel® Architecture with version 2.8
What's New in AI Analytics Toolkit 2023.1
-
Accelerate your AI training and inference software performance using the 2023.1 AI Kit. In this release, Intel® Xe Matrix Extensions (Intel® XMX) in Intel Flex and Max Series GPUs are exposed to developers by Intel® oneAPI Deep Neural Network Library (oneDNN), through deep learning frameworks such as TensorFlow and PyTorch, for increased performance across wide range of market segments to deliver competitive performance.
- Additional significant performance gains are provided by Intel® Extensions for TensorFlow (ITEX) and Intel® Extensions for PyTorch (IPEX), both having native GPU support.
System Requirements
Please see system requirements.
Please reference the usage guides for each of the included tools:
- Intel® Optimization for TensorFlow* - See Performance Considerations
- Intel® Optimization for PyTorch * - See Performance Considerations
Known Limitation
- This release of Model Zoo does not support workloads on Intel® Data Center GPU Max Series, but supports workloads on Intel® Data Center GPU Flex Series.
- Intel® AI Analytics Toolkit (AI Kit) 2023.1 doesn't work with Intel® oneAPI Base Toolkit 2023.1 because of incompatibility with Intel® oneAPI DPC++/C++ Compiler and runtime and Intel® oneAPI Math Kernel Library. Please keep using Intel® oneAPI Base Toolkit 2023.0 with Intel AI Analytics Toolkit 2023.1.
-
Intel® Optimization for TensorFlow* is compatible with version 2.9.1
-
Intel® Optimization for PyTorch* is compatible with version 1.12.1
-
Intel® Distribution for Python* is compatible with cpython version 3.9
-
Intel® Distribution of Modin* is compatible with version 0.17
-
Intel® Neural Compressor is compatible with version 1.13
-
Model Zoo for Intel® Architecture with version 2.8
What's New in AI Analytics Toolkit 2023.0
- Easily scale your AI solution from low compute resources to large or distributed compute resources with the new Heterogeneous data kernals (HDK) backend for Modin.
- Use Accenture reference kits to jumpstart your AI solution.
- Choose your platform, as AI Kit now supports running natively on Windows with full parity to Linux (except for distributed training).
- Bzip2 is required to extract packages, and a Linux system without this software will result in an installation failure.
- Intel® Optimization for TensorFlow* compatible with version 2.9.1 may not include all the latest functional and security updates. A new version of Intel® Optimization for TensorFlow* is targeted to be released by March 2023 and will include additional functional and security updates. Customers should update to the latest version as it becomes available.
Compatibility Notes
-
Intel® Optimization for TensorFlow* is compatible with version 2.9.1
-
Intel® Optimization for PyTorch* is compatible with version 1.12.1
-
Intel® Distribution for Python* is compatible with cpython version 3.9
-
Intel® Distribution of Modin* is compatible with version 0.13.3
-
Intel® Neural Compressor is compatible with version 1.13
-
Model Zoo for Intel® Architecture with version 2.8
What's New in AI Analytics Toolkit 2022.3
-
Product Review Windows OS support
-
Intel® Optimization for TensorFlow*
- Updated to TensorFlow v2.9.1, and compiler options are no longer needed to enable the current oneDNN v2.6 optimizations on Intel Cascade lake and newer CPUs on Linux OS
- Tensorflow now includes performance improvements for Bfloat16 models with AMX optimizations and more operations are supported with BF16 datatype
- CVE and bug fixesl
-
Intel® Optimization for PyTorch*
- Intel® Extension for PyTorch has been updated to 1.12.100, following PyTorch v1.12.
- PyTorch now includes automatic int8 quantization and made it a stable feature. Runtime extension is stabilized and MultiStreamModule feature is brought to further boost throughput in offline inference scenario.
- Enhancements in operation and graph has been added which are positive for performance of broad set of workloads.
- Intel® Neural Compressor
- Intel® Neural Compressor mainly updated with Tensorflow new quantization API support, QDQ quantization support for ITEX, mixed precision enhancement, DyNAS support, training for block-wise structure sparsity support, op-type wise tuning strategy support.
- Intel® Neural Compressor improved productivity with lighter binary size, quantization accuracy diagnostic feature support by GUI, and experimental auto-coding support.
-
Intel® Model Zoo
- Model Zoo for Intel® Architecture now supports Intel® Optimization for PyTorch v1.12.1 and Intel® Optimization for TensorFlow v2.9.1, while also including fixes and improvements for product quality and use of current stable releases. Experimental support is added for SLES and SUSE platforms.
Compatibility Notes
-
Intel® Optimization for TensorFlow* is compatible with version 2.8
-
Intel® Optimization for PyTorch* is compatible with version 1.10
-
Intel® Distribution for Python* is compatible with cpython version 3.9
-
Intel® Distribution of Modin* is compatible with version 0.13
-
Intel® Neural Compressor is compatible with version 1.10.1
What's New
-
Product Review Windows OS support
-
Intel® Optimization for TensorFlow*
-
Performance improvements for Bfloat16 models with AMX optimizations.
-
Enabled support for 12th Gen Intel(R) Core (TM) (code named Alder Lake) platform.
-
No longer supports oneDNN block format, i.e., setting TF_ENABLE_MKL_NATIVE_FORMAT=0 will not enable blocked format.
-
To enable AMX optimization, you no longer need DNNL_MAX_CPU_ISA = AVX512_CORE_AMX.
-
Updated oneDNN to version 2.5.1
-
Improved _FusedMatMul operation, which enhances the performance of models like BERT
-
Added LayerNormalization ops fusion and BatchMatMul – Mul – AddV2 fusion to improve performance of Transformer based language models
-
Improved performance of EfficentNet and EfficientDet models with addition of swish (Sigmoid – Mul) fusion
-
Removed unnecessary transpose elimination to enhance performance for 3DUnet model
-
-
Intel® Optimization for PyTorch*
- changed the underhood device from XPU to CPU. With this change, the model and tensor do not need to be converted to the extension device to get a performance improvement.
- optimize the Transformer* and CNN models by fusing more operators and applying NHWC.
- Change the package name to intel_extension_for_pytorch while the original package name is intel_pytorch_extension.
- support auto-mixed-precision for Bfloat16 data type.
- provides the INT8 calibration as an experimental feature while it only supports post-training static quantization now.
- Intel® Neural Compressor
- Supported the quantization on latest deep learning frameworks
- Supported the quantization for a new model domain (Audio)
- Supported the compatible quantization recipes for framework upgrade
- Supported fine-tuning and quantization using INC & Optimum for “Prune Once for All: Sparse Pre-Trained Language Models” published at ENLSP NeurIPS Workshop 2021
- Proved the sparsity training recipes across multiple model domains (CV, NLP, and Recommendation System)
- Improved INC GUI for easy quantization
- Supported quantization on 300 random models
- Added bare-metal examples for Bert-mini and DLRM
-
Intel® Model Zoo
- Add support for TensorFlow v2.7.0
- Support for PyTorch v1.10.0 and IPEX v1.10.0
- Added new Diagnostics Utility for Intel® oneAPI Toolkits to diagnose the system status for using Intel® products. Learn more.
-
Intel® Optimization for TensorFlow* is compatible with version 2.6
-
Intel® Optimization for PyTorch* is compatible with version 1.8
-
Intel® Distribution for Python* is compatible with cpython version 3.9
-
Intel® Distribution of Modin* is compatible with version 0.8.2
-
Intel® Neural Compressor is compatible with version 1.7
-
Intel oneAPI Intel® AI Analytics Toolkit 2022.1.2 has been updated to include functional and security updates including Apache Log4j* version 2.17.1. Users should update to the latest version as it becomes available.
-
Intel® Optimization for PyTorch*
- The Intel® Extension for PyTorch now supports Python 3.9 and Microsoft's Windows Subsystem for Linux (WSL).
- Intel® Neural Compressor
- Supported magnitude pruning on TensorFlow
-
Supported knowledge distillation on PyTorch
-
Supported multi-node pruning with distributed dataloader on PyTorch
-
Supported multi-node inference for benchmark on PyTorch
-
Improved quantization accuracy in SSD-Reset34 and MobileNet v3 on TensorFlow
- Supported the configuration-free (pure Python) quantization
- Improved ease-of-use user interface for quantization with few clicks
-
Added a domain-specific acceleration library for NLP models
-
Intel® Model Zoo
- Add support for TensorFlow v2.6.0 and TensorFlow Serving v2.6.0
- Add support for Pytorch 1.9 and Intel Extension for Pytorch 1.9
- Add Transfer Learning sample IPython Notebook that fine-tunes BERT with the IMDb dataset
- Add documentation to create an Intel Neural Compressor container with Intel Optimized TensorFlow
- Additional models and precisions:
- ML-Perf Transformer-LT Training (FP32 and BFloat16)
- ML-Perf Transformer-LT Inference (FP32, BFloat16 and INT8)
- ML-Perf 3D-Unet Inference (FP32, BFloat16 and INT8)
- DIEN Training (FP32)
- DIEN Inference (FP32 and BFloat16)
-
Intel® Distribution for Python*
-
Intel® Distribution for Python now supports Python version 3.9
-
The dpctl package offers developers increased debugging capabilities with improved error handing and reporting
- Data Parallel Python technology now provides zero copy data exchange performance across packages
-
- Added new Diagnostics Utility for Intel® oneAPI Toolkits to diagnose the system status for using Intel® products. Learn more.
Known Issue
- Compatibility issue between Intel® AI Analytics Toolkit and Intel® Base Analytics Toolkit.
- Intel Distribution for Python in Basekit 2022.2 Release updated three packages cryptography / pyopenssl / libprotobuf causing package confliction with TensorFlow in AI Kit 2022.1 Release.
- Solution : Install AI Kit 2022.1 Release in a separate directory.
sudo ./l_AIKit_b_2021.1.8.618_offline.sh -s -a --install-dir /target/install/path --silent --eula accept
-
Intel® Optimization for TensorFlow* is compatible with version 2.5
-
Intel® Optimization for PyTorch * is compatible with version 1.8
-
Intel® Distribution for Python* is compatible with cpython version 3.7
-
Intel® Distribution of Modin* is compatible with version 0.8.2
-
Intel® Low Precision Optimization Tool is compatible with version 1.5.1
-
Fine tune the performance of Natural Language algorithms through the latest sparsity and pruning features introduced in AI Analytics Toolkit.
- Intel® Low Precision Optimization Tool
-
EGradient-sensitivity pruning for CNN model
-
Static quantization support for ONNX NLP model
-
Dynamic seq length support in NLP dataloader
-
Enrich quantization statistics
-
-
Intel® Model Zoo
-
Added links to Intel oneContainer Portal
-
Added documentation for running most workflows inside Intel® oneAPI AI Analytics Toolkit
-
Experimental support for running workflows onCentOS 8
-
-
Intel® Distribution for Python*
-
Intel® Distribution for Python* release notes
-
Please see system requirements.
Please reference the usage guides for each of the included tools:
- Intel® Optimization for TensorFlow* - See Performance Considerations
- Intel® Optimization for PyTorch * - See Performance Considerations
- Intel® Optimization for TensorFlow*
- Int8 will only work when environment variable TF_ENABLE_MKL_NATIVE_FORMAT=0 is set.
Compatibility Notes
-
Intel® Optimization for TensorFlow* is compatible with version 2.5
-
Intel® Optimization for PyTorch * is compatible with version 1.8
-
Intel® Distribution for Python* is compatible with cpython version 3.7
-
Intel® Distribution of Modin* is compatible with version 0.8.2
-
Intel® Low Precision Optimization Tool is compatible with version 1.4.1
What's New
-
Intel® Optimization for TensorFlow*
- oneDNN - upgraded oneDNN to v2.2
- The default for Intel TF is now native format, The user will need to set the env-variable TF_ENABLE_MKL_NATIVE_FORMAT=0 to use blocked formats.
- OneDNN primitive cache enabled. Improved performance of models with batch size 1.
- Various ops fusions with FP32, BFloat16, and INT8 data format
- Conv2D+Squeeze+BiasAdd Fusion
- MatMul+BiasAdd+Add Fusion.
- Enabled MatMul + Bias + LeakyRelu Fusion.
- CNMS performance optimization
- Enabled DNNL CPU dispatch control.
- Graph pattern match for grappler op fusion optimization
- Supporting quantized pooling op for signed 8 bits.
- Enable MklConv/MklFusedConv with explicit padding
- Remove nGraph build support tensorflow#42870
- Execute small gemm's single threaded.
- Removed unnecessary OneDNN dependencies.
- Removed DNNL 0.x support
- Bug fixes
- Issues resolved in TensorFlow 2.5
- oneDNN resolved issues. 2.2 resolved issues
- Fixed memory leak in MKLAddN
- Fixed the bug to duplicate kernel registration of BatchMatMulV2.
- Fixed unit test failures due to benchmark test API changes
- incorrect result of _MKLMaxPoolGrad 40122.
- Intel® Optimization for PyTorch *
- Upgraded the oneDNN from v1.5-rc to v1.8.1
- Updated the README file to add the sections to introduce supported customized operators, supported fusion patterns, tutorials and joint blogs with stakeholders
- Intel® Model Zoo
-
One new PyTorch workload containers and model packages that are available on the Intel® oneContainer Portal:
- DLRM BFloat16 Training
-
Two new TF workload containers and model packages that are available on the Intel® oneContainer Portal:
- 3D U-Net FP32 Inference
- SSD ResNet34 BFloat16 Training
-
- Intel® Low Precision Optimization Tool
-
Extended Capabilities
- Model conversion from QAT to Intel Optimized TensorFlow model
-
User Experience
- More comprehensive logging message
- UI enhancement with FP32 optimization, auto-mixed precision (BF16/FP32), and graph visualization
- Online document
-
Model Zoo
- INT8 key models updated (BERT on TensorFlow, DLRM on PyTorch, etc.)
- 20+ HuggingFace model quantization
-
Pruning
- Pruning/sparsity API refinement
- Magnitude-based pruning on PyTorch
- Quantization
- PyTorch FX-based quantization support
- TensorFlow & ONNX RT quantization enhancement
-
- Intel® Distribution for Python*
- Intel® Distribution for Python* release notes
Known Limitation
- Intel® Optimization for TensorFlow*
- Int8 will only work when environment variable TF_ENABLE_MKL_NATIVE_FORMAT=0 is set.
Compatibility Notes
-
Intel® Optimization for TensorFlow* is compatible with version 2.3
-
Intel® Optimization for PyTorch * is compatible with version 1.7
-
Intel® Distribution for Python* is compatible with cpython version 3.7
-
Intel® Distribution of Modin* is compatible with version 0.8.2
-
Intel® Optimization for TensorFlow*
-
oneDNN - moved from 0.x to version 1.4
- Support for Intel® MKL-DNN version 0.x is still available.
- Issues resolved in TensorFlow 2.3
- oneDNN resolved issues
-
-
Intel® Optimization for PyTorch *
-
New PyTorch * 1.7.0 was newly supported by Intel extension for PyTorch *.
-
Device name was changed from DPCPP to XPU.
-
Enabled the launcher for end users.
-
Improvement for INT8 optimization with refined auto mixed precision API.
-
More operators are optimized for the int8 inference and bfp16 training of some key workloads, like: MaskRCNN, SSD-ResNet34, DLRM, RNNT.
-
New custom operators: ROIAlign, RNN, FrozenBatchNorm, nms.
-
Performance improvement for several operators: tanh, log_softmax, upsample, embeddingbad and enables int8 linear fusion.
-
Bug fixes
-
-
Intel® Model Zoo
- Several new TensorFlow* and PyTorch* models added to the Intel® Model Zoo.
- Ten new TensorFlow workload containers and model packages that are available on the Intel® oneContainer Portal
- Two new PyTorch workload containers and model packages that are available on the Intel® oneContainer Portal
- Three new TensorFlow Kubernetes packages that are available on the Intel® oneContainer Portal:
- A new Helm chart to deploy TensorFlow Serving on a K8s cluster
- Bug-fixes, improvements to documentations
-
Intel® Low Precision Optimization Tool
- New backends (PyTorch/IPEX, ONNX Runtime) preview support
- Add built-in industry dataset/metric and custom registration
- Preliminary input/output node auto-detection on TensorFlow models
- New INT8 quantization recipes: bias correction and label balance
- 30+ OOB models validated
- 2021.2 AI Kit installation causes a Intel® oneapi Base Toolkit 2021.3 installation issue if users don't have 2021.2 Base Toolkit installation alongside the 2021.2 AI Kit installation.
- To avoid this issue, Users could install Intel® oneapi Base Toolkit 2021.2 alongside the 2021.2 AI Kit installation first, and then Intel® oneapi Base Toolkit 2021.3 could be installed successfully.
- Intel® Optimization for PyTorch*
- Multi-node training still encounter hang issues after several iterations. The fix will be included in the next official release.
Compatibility Notes
-
Intel® Optimization for TensorFlow* is compatible with version 2.2
-
Intel® Optimization for PyTorch * is compatible with version 1.5
-
Intel® Distribution for Python* is compatible with cpython version 3.7
-
Intel® Distribution of Modin* is compatible with version 0.8.2
Known Limitation
Intel® Optimization for PyTorch* requires Intel® AVX512 supports on the system.
Notices and Disclaimers
Intel technologies may require enabled hardware, software or service activation.
No product or component can be absolutely secure.
Your costs and results may vary.
© Intel Corporation. Intel, the Intel logo, and other Intel marks are trademarks of Intel Corporation or its subsidiaries. Other names and brands may be claimed as the property of others.
No license (express or implied, by estoppel or otherwise) to any intellectual property rights is granted by this document.
The products described may contain design defects or errors known as errata which may cause the product to deviate from published specifications. Current characterized errata are available on request.
Intel disclaims all express and implied warranties, including without limitation, the implied warranties of merchantability, fitness for a particular purpose, and non-infringement, as well as any warranty arising from course of performance, course of dealing, or usage in trade.