On Wed, Sep 8, 2021 at 3:43 AM Linus Torvalds torvalds@linux-foundation.org wrote:
On Tue, Sep 7, 2021 at 6:35 PM Linus Torvalds torvalds@linux-foundation.org wrote:
I think a lot of them have just copied the x86 code (it was 4k long ago), without actually understanding all the details.
Just to put the x86 number in perspective: it was raised to 8192 back in 2013, with the comment
x86/cpu: Increase max CPU count to 8192 The MAXSMP option is intended to enable silly large numbers of CPUs for testing purposes. The current value of 4096 isn't very silly any longer as there are actual SGI machines that approach 6096 CPUs when taking HT into account. Increase the value to a nice round 8192 to account for this and allow for short term future increases.so on the x86 side, people have actually done these things.
Other architectures? I think some IBM power9 machines can hit 192 cores (with SMT4 - so NR_CPUS of 768), but I don't think there's been an equivalent of an SGI for anything but x86.
But admittedly I haven't checked or followed those things. I could easily imagine some boutique super-beefy setup.
POWER10 was just announced with threads 1920 using SMT8, I think the latest s390 and sparc64 (from 2017) are in the same ballpark when using SMT. The largest arm64 I know of was ThunderX3 with 768 threads on dual-socket machines. This got cancelled before it was shipped to customers, but it's likely that others will exceed that in the future.
Arnd