Post #1396187
2026-03-31 11:18 UTC
@fmarini
With Apple's design, including very close integration of CPU and GPU talking to the same RAM on the same package? No, I do not think so.
Most of the GPUs wouldn't be connected to screens rendering them useless. They are not huge complex nVidia style GPUs able to do GPGPU stuff.
Therefore they are a waste of silicon, meaning money, meaning yield... even if they're turned off and not making heat.
No. I don't buy it.
That does not of course mean that other huge-scale integration based around Arm cores couldn't be done. Which is exactly what the Ampere and Graviton chips are doing... even if they have basically just a single customer: Amazon.
Replies (1)
-
@fmarini@mastodon.social 2026-03-31 12:11
@lproven sorry, but the AS GPUs are absolutely capable of doing general purpose computation. What is missing is CUDA support, but that doesn’t absolutely mean they’re not more than good at GPGPU. The fact that the RAM is local to each AS processor is why I wrote NUMA system. Moreover, every multi-GPU system is de facto a NUMA one, since the graphics memory is local to each single GPU card, and each GPU can only talk to its memory. Also, they’re not connect to any output, usually.