What it does

nGraph was Intel’s attempt to build a universal middle layer for deep learning: write your model in TensorFlow, ONNX, or another framework, and nGraph would compile it down to run on CPUs, GPUs, or Intel’s now-canceled Nervana NNP accelerators. It shipped as a Python wheel and promised performance boosts through subgraph pattern matching.

The interesting bit

The README now functions as a forwarding address. Every page leads to the same punchline: “nGraph has moved to OpenVINO.” The project still advertises a 45× performance claim and support for hardware (NNP-T, NNP-I) that Intel killed years ago, which gives the repo a distinctly ghost-ship quality.

Key highlights

• Supported TensorFlow, ONNX, MXNet, PyTorch, and PaddlePaddle as frontends
• Integrated with PlaidML for GPU acceleration across Intel, NVIDIA, and AMD cards
• Provided validated workload lists (20 for TensorFlow, 17 for ONNX at the time)
• Apache 2.0 licensed; build system used standard make check for CI
• Now lives entirely inside OpenVINO; this repo is archived in place

Caveats

• The “up to 45X” performance claim links to a 2019 Intel blog post; no reproducible benchmarks are in the repo itself
• macOS support stopped at Mojave (10.14.3), and pip requirements pin to versions from 2019
• NNP-T and NNP-I hardware references are historical artifacts; Intel discontinued the Nervana line

Verdict

Worth a look if you’re tracing the genealogy of OpenVINO or studying how ML compilers evolved. Active developers should head straight to the OpenVINO repository instead; this one is a well-preserved signpost, not a destination.