AMD hands ZenDNN over to Intel: AI on CPUs finally no longer fighting separate battles
AMD folds ZenDNN into Intel: AI on CPUs finally stops fighting separate battles. AMD and Intel, these two sworn rivals who have been at each other's throats for 50 years, have shaken hands on AI—AMD announced it is merging its own CPU AI acceleration library into Intel-led oneDNN, so that from now on, when running AI inference on CPUs, the two companies' chips will follow a single code path. First, let's talk about what was done.
💡 What You Will Learn
AMD folds ZenDNN into Intel: AI on CPUs finally stops fighting separate battles. AMD and Intel, these two sworn rivals who have been at each other's throats for 50 years, have shaken hands on AI—AMD a
📜 Table of Contents
AMD Merges ZenDNN into Intel: AI on CPUs Finally Stops Fighting Separate Battles
AMD and Intel, rivals for 50 years, have shaken hands on AI—AMD announced it's merging its CPU AI acceleration library into Intel-led oneDNN, so from now on, running AI inference on CPUs will follow a single code path across both vendors' chips.
First, what happened. AMD had an acceleration library called ZenDNN, a downstream oneDNN branch that AMD maintained itself, specifically to accelerate PyTorch (the mainstream AI framework) on AMD chips. This time AMD is merging ZenDNN entirely back into oneDNN upstream, and Intel has agreed. After the merge, oneDNN is a single library that tunes for both AMD and Intel chips—developers no longer need to configure separate optimizations for each vendor, and switching platforms doesn't require rewriting.
This move was brokered by the x86 Ecosystem Advisory Group (EAG), a coordination body founded in 2024 with members from both sides. They'd already collaborated on the ACE instruction set, so this is essentially following the same thread to straighten out the AI software path on CPUs.
One library, speeding up AI on CPUs (1.24x and 2.21x)
oneDNN (an open-source deep learning acceleration library) automatically detects hardware at runtime, generates optimized kernels on the spot for the current chip, and supports low-precision data types, with targeted acceleration for high-frequency AI operations like matrix multiplication and convolution. When it was integrated into PyTorch in late 2023, Float32 inference got a 1.24x speedup and BFloat16 got 2.21x. And it's not just x86—Arm and RISC-V are covered too—its lineage traces back to Intel's early deep learning library, later merged into the UXL Foundation (the open-source organization that hosts oneDNN).
AMD Pays to Get Upstream (US$20,000 buys a voice)
This time AMD isn't just merging code—it's also announced joining the UXL Foundation as a steering member, at US$20,000 a year, in exchange for a voice in the foundation's direction. From maintaining its own branch to paying to get upstream and set the rules together, this posture says more than the code itself: both companies are genuinely playing this as a long game.
What it means for users (CPU inference is underestimated)
For developers running PyTorch, this is real convenience. For ordinary people, here's the other thing I want to say: AI inference on CPUs is severely underestimated. Not all AI runs on graphics cards—CPU inference on servers and in data centers is the bulk of it, and data center CPU demand is being pushed up by AI agents. This unification lowers the barrier to running AI on CPUs another notch.
The deepest card buried here is ACE (the future of matrix multiplication)
Merging ZenDNN is only the first layer. The two companies have long been collaborating on ACE (AI Compute Extensions), a next-generation x86 instruction set specifically for accelerating matrix multiplication—the core of AI computation. The plan is for ACE to also be exposed to PyTorch through oneDNN in the future. With both chip instructions and software libraries pushing forward together, x86 has laid the road to shed its old image of only being able to rely on graphics cards in the AI era.
Developers used to have to pick a side between AMD and Intel; now the code is the same, so either works. Is your machine AMD or Intel? When you run local AI, who's doing the work—the CPU or the graphics card? Let's talk in the comments.
Written by our editorial team; tools listed here are tested or verified against public sources. Links point to official sites or GitHub repos for reference only — no paid placements.