The June launch of Windows Copilot+ PCs has come and gone. Support for developers is not yet at the level it should be, but Microsoft is moving in the right direction with DirectML.
From now on, the DirectML library also works on the Hexagon NPU inside the Snapdragon X Elite that powers Arm-based Windows laptops. DirectML is to machine learning applications what DirectX has long been to graphics solutions: an abstraction layer to easily accelerate specific workloads through hardware.
DirectML expansion
DirectML already worked on Intel and AMD’s x86-based NPUs. With DirectML’s API layer now expanded to the Qualcomm offering, it looks like developers can optimize their AI tasks for Arm without too much trouble. The much-discussed TOPS number, the favoured yardstick for pm=devoce AI performance, will soon be very similar for Intel, AMD and Qualcomm. Once Intel launches Lunar Lake, all three chip companies will be above the 40 TOPS benchmark Microsoft deems necessary for a Copilot+ PC.
These specifications will eventually run software such as Windows Recall. This AI feature allows users to retrieve snapshots from their own PC usage. Although Recall will be reintroduced soon, it was immediately controversial because of the privacy and security implications of a computer recording and saving your screen every few seconds.
Support for developers?
Any announcement about a further rollout of Arm-based Windows support is welcome. Since Microsoft and Qualcomm made a big splash around the launch of Copilot+ PCs in June, the attitude of both parties has been anything but proactive. For example, things have been quiet for several months around dev kits for Snapdragon X Elite systems. Developers are struggling to get their hands on them.
The severity of this problem varies. After all, x86 apps do in fact still run on Snapdragon chips apart from various video games and some niche solutions. Prism, the emulation layer that runs x86 applications on Arm-based Windows PCs is claimed to be as good as Apple’s famous Rosetta. The latter enabled Apple to move fairly gracefully from Intel’s x86 chips to its own Arm-based M processors. Rosetta was really a powerhouse back then, too: non-optimized applications still ran faster on Apple Silicon than on Intel hardware in Macs that had come out months earlier.
The same is not true of Prism. XDA showed the performance loss during a benchmark of the AI Denoise feature in Adobe Lightroom Classic. Whereas Intel x86 chips took anywhere from half a minute to just under four minutes, the Snapdragon X Elite only completed the test after ten (!) minutes. Things like native DirectML support are thus really needed.
Also read: Smartphones vulnerable due to Qualcomm GPU drivers, Google discovers