PlaidML

PlaidML
Name	PlaidML
Developer	Intel Corporation
Initial release	2018
Programming language	Python, C++
Operating system	Cross-platform
License	Apache License 2.0

PlaidML PlaidML is a machine learning compiler and tensor compiler designed to provide portable deep learning acceleration across diverse hardware. It was created to enable frameworks and researchers to run neural network models on GPUs and accelerators from multiple vendors, integrating with projects and institutions in the broader open source ecosystem. The project has interfaced with major frameworks and toolchains while evolving through corporate stewardship and community contributions.

Overview

PlaidML was developed to address portability challenges between hardware vendors such as NVIDIA, AMD, and Intel Corporation while supporting frameworks including TensorFlow, Keras (software), and ONNX. The project emphasized a retargetable intermediate representation and code generation approach influenced by research from institutions like Berkeley AI Research, Massachusetts Institute of Technology, and Stanford University. Its goals intersected with initiatives from organizations including Xilinx, ARM Limited, and Google (company) to broaden access to GPU acceleration for developers affiliated with projects such as PyTorch and Apache MXNet. PlaidML’s design choices reflect needs expressed in conferences and venues such as NeurIPS, ICLR, and ACM SIGGRAPH for efficient deployment across heterogeneous systems like servers from Dell Technologies, workstations from Apple Inc., and edge devices from Raspberry Pi partners.

Architecture and Design

The architecture centers on a tensor expression language and a compiler stack inspired by academic work at University of California, Berkeley and Carnegie Mellon University. A core scheduler and optimizer translate computational graphs from frontends such as Keras (software), TensorFlow, and ONNX into low-level code for backends provided by vendor ecosystems like OpenCL, Vulkan (API), and driver stacks from NVIDIA Corporation. The design separates concerns between graph lowering, memory planning, and code generation, following methodologies practiced in projects like LLVM, Halide (software), and TVM (deep learning) to enable autotuning similar to research from Intel Labs and Facebook. Components include a frontend adapter layer compatible with tools deployed by teams at Google Brain, OpenAI, and DeepMind and a backend code generator that can emit kernels compatible with runtime systems influenced by Mesa (computer graphics), ROCm, and CUDA ecosystems.

Supported Platforms and Backends

PlaidML targeted a wide set of execution targets including discrete GPUs from NVIDIA Corporation and AMD, integrated GPUs from Intel Corporation, and mobile GPUs from Qualcomm. It used compute APIs such as OpenCL, Vulkan (API), and vendor runtimes influenced by CUDA and ROCm to run workloads on systems marketed by Lenovo, HP Inc., and Apple Inc.. The project supported operating environments like Linux, Windows, and macOS and was tested on hardware families from GeForce (brand), Radeon (brand), and Intel Iris Graphics. Integration points were created for cloud providers and services such as Amazon Web Services, Microsoft Azure, and Google Cloud Platform where accelerators from NVIDIA Tesla, AMD Instinct, and Intel Xe are deployed.

Performance and Benchmarks

Performance characteristics were evaluated against frameworks and toolchains including TensorFlow, PyTorch, and MXNet using models like ResNet (neural network), MobileNet, and BERT (language model). Benchmarks published by contributors compared kernel throughput, memory utilization, and end-to-end latency on hardware from NVIDIA Corporation and AMD against established toolchains like cuDNN and oneDNN. Autotuning strategies borrowed from research at University of Illinois Urbana–Champaign and labs at Facebook AI Research sought to close gaps in throughput reported in community benchmarks presented at venues like MLSys and ISCA. Real-world comparisons often highlighted trade-offs in driver maturity, API overhead, and kernel fusion strategies when juxtaposed with vendor-optimized libraries from NVIDIA and Intel.

Integration and Ecosystem

PlaidML provided adapters and bindings for frontends such as Keras (software), TensorFlow, and ONNX to enable model import and execution in projects affiliated with developers from GitHub, Google, and Microsoft Research. Its integration footprint touched tooling in the Python ecosystem including package managers used by contributors from Anaconda, Inc., continuous integration systems like Jenkins, and container orchestration platforms such as Kubernetes. The open source community around the project included contributors affiliated with companies like Intel Corporation, research labs such as Berkeley AI Research, and independent developers active on platforms like GitHub and in conferences like FOSDEM and Open Source Summit.

Development History and Maintenance

The project originated in the late 2010s with community and corporate involvement from entities including Intel Corporation and contributors known from repositories on GitHub. Development activity tracked upstream changes influenced by research from Stanford University and UC Berkeley and by engineering practices practiced at Google, Facebook, and Microsoft. Maintenance patterns included releases, issue triage, and contributions coordinated through tools and services used by organizations such as GitHub, Travis CI, and CircleCI. Over time stewardship shifted as corporate priorities and open source strategies evolved, paralleling transitions seen in other projects associated with Xilinx, ARM Limited, and academic incubations at MIT.

Category:Deep learning software