Ampere Computing just launched a groundbreaking server CPU boasting an impressive 192 processor cores. But there’s a hiccup when trying to use these in a single system with a dual-socket motherboard. You see, the standard ARM-64 kernel for Linux can only handle up to 256 cores, which becomes a snag in this setup. To tackle this issue, Ampere has submitted a patch aiming to lift this limitation.
Their solution involves tapping into a method called CPUMASK_OFFSTACK. This nifty trick allows for the allocation of free bitmaps from memory to handle CPU masks. The idea is to empower the Linux kernel to smoothly manage up to 4096 cores without bulking up the kernel images, which are initially set at 8 kilobytes per CPU core.
Until now, 256 cores sufficed for most needs, especially in supercomputers or systems with multiple CPUs. But in this specific case, the kernel needs tweaking via CPUMASK_OFFSTACK for a higher core limit.
For instance, AMD’s top-tier server CPU offers 128 Zen-4c cores, making it a breeze to operate in a dual-socket system. On the other hand, the Intel Xeon Platinum 8480+ packs 56 cores, needing kernel recompilation only if you exceed four CPUs per system.
This workaround seems to be holding up for the time being
It might be a while before the standard configuration of the Linux kernel caters to CPUs surpassing the 256-core mark. There was word that a patch addressing this was turned down in 2021 due to the lack of ARM-based hardware surpassing the existing limit at the time.
Consequently, the upcoming Linux kernel version 6.8 might still rely on the optional CPU-Mask-Offstack method to keep the kernel image size in check.
Shifting gears to the desktop arena, Microsoft might find itself needing adjustments down the road. At present, Windows 11 Home and Pro support only up to 64 threads. Beyond that, extra cores get recognized but allocated to a new core group.
Most software isn’t equipped to divvy up its workload across these different core groups. An AMD Ryzen Threadripper, boasting 48 cores, already pushes past the limit of what regular Windows versions can handle.
