4 ms·
> The M1 chips have a unified memory architecture shared between GPU and CPUs. This is a masterstroke for performance. No it isn't. It's a nice optimization to
by kllrnohj 3y ago
> The M1 chips have a unified memory architecture shared between GPU and CPUs. This is a masterstroke for performance.
No it isn't. It's a nice optimization to have, but it doesn't radically change performance. If unified memory was that drastic AMD's APUs from 2012 to 2015 wouldn't have been so forgettable which is where heterogenous compute with unified GPU/CPU memory was pioneered.
Also every phone SoC has unified memory, and yet only Apple's can compete with Intel/AMD x86 on performance. M1's performance ain't from unified memory (especially since mixed CPU/GPU compute on the same data is very rare)
- zeusk 3y agoAMD's APUs or the Windows Display Driver Model did not support USM until very recently. They would explicitly carve out memory at boot for graphics use and physical pages could not be shared between GPU and CPU. They'd still make use of apertures to copy data in/out of respective memory ranges.
- kllrnohj 3y agoThat is simply not true. Kavari (2014) had pointer passing between CPU & GPU. I don't know where you're getting that pages couldn't be shared from but it's not accurate. https://images.anandtech.com/doci/7677/03%20-%20HSA%20Features.jpg https://images.anandtech.com/doci/7677/03%20-%20HSA%20Featur... https://images.anandtech.com/doci/7677/17%20-%20hUMA%20Benefit.JPG https://images.anandtech.com/doci/7677/17%20-%20hUMA%20Benef... https://www.anandtech.com/show/7677/amd-kaveri-review-a8-7600-a10-7850k/6 https://www.anandtech.com/show/7677/amd-kaveri-review-a8-760...
- zeusk 3y agofor compute only (restricted to openCL), not graphics workloads.