intel
Last month, Microsoft declared 2024 the “Year of the AI PC” when it announced the new Copilot key for PC keyboards. On one level, this is just an aspirational PR-friendly declaration, signaling to investors that Microsoft intends to continue driving the AI hype cycle as it competes with Apple for the title of most valuable publicly traded company. It is intended to show.
But on a technical level, it’s true that PCs manufactured and sold after 2024 will generally have AI and machine learning processing capabilities that older PCs don’t have. The main one is the Neural Processing Unit (NPU). This is a special block found in recent high-end Intel and AMD CPUs that allows certain generative AI and machine learning workloads to run on the CPU or GPU.
Qualcomm’s Windows PCs were some of the first PCs to feature NPUs, and the Arm processors used in most smartphones have had some form of machine learning acceleration for several years (Apple’s Mac All M-series chips also feature an NPU). dating back to his M1 in 2020). But the Arm version of Windows is a small, insignificant part of the overall PC market. x86 PCs with Intel’s Core Ultra chips, AMD’s Ryzen 7040/8040 series laptop CPUs, or Ryzen 8000G desktop CPUs will be many mainstream PC users’ first exposure to this type of hardware.
Right now, even if your PC has an NPU, Windows can’t use it much except for webcam background blur and a few other video effects. But that’s slowly changing, and part of that is making it relatively easy for developers to create his NPU-independent apps in the same way that PC game developers currently create GPU-independent games. You will be able to create one.
A gaming example is instructive. Because this is essentially how Microsoft approaches DirectML, its API for machine learning operations. Until now, these AI workloads have been primarily used to run on GPUs, but last week Microsoft announced that DirectML support for Intel’s Meteor Lake NPUs will be available in developer preview starting with DirectML 1.13.1 and ONNX Runtime 1.17. announced that it would add.
Runs only an unspecified “subset of supported machine learning models”, with some models “may not run at all or may have high latency or low accuracy”, but more Opens the door to third-party apps. Start taking advantage of the built-in NPU. Intel says Samsung uses Intel’s NPU and DirectML for facial recognition in its Photo Gallery app, and Apple also uses its Neural Engine in macOS and iOS.
This provides significant benefits compared to running these workloads on GPUs or CPUs.
Robert Hallock, Intel’s senior director of technical marketing, said in an interview with Ars about Meteor Lake’s capabilities, “At least in the Intel space, NPUs will be used primarily for power efficiency reasons.” “Camera segmentation, background blur, all that stuff… moving this to the NPU saves about 30-50% power compared to running it elsewhere.”
Both Intel and Microsoft are working towards a model where NPUs are treated much like current GPUs. Developers typically target DirectX rather than a specific graphics card manufacturer or his GPU architecture, and new features, one-time bug fixes, and performance improvements are all possible. This may be addressed by updating the GPU driver. Some GPUs run certain games better than others, and a developer may choose to spend more time optimizing his Nvidia or AMD card, but in general the model is does not depend on
Similarly, Intel already provides GPU-style driver updates for NPUs. Also, Hallock says that Windows basically already recognizes NPUs as what he calls “graphics cards without rendering capabilities.”
