Neural Engine outperforming the GPU on the M1 Max
Photography is my hobby, and I’ve not bought a new computer since 2016, so the new Macbook Pro was my choice, I went for the M1 Max processor because it will also be my dev box for projects, alongside the Raspberry Pi 4s.
Here is an interesting thing though, Photoshop for M1 was faster turning a set of 20 photos into a panorama, taking 2:17 against the old 2016 MacBook Pro (with the Radeon GPU), which took 6:51. Pretty much 3x, which is great…
But 3x doesn’t really seem that much TBH for kit that is 5 years old, Moore’s Law would have it working at 5x.
However, then I opened up DXO Photolab which I like to use for RAW processing. Now DXO isn’t native M1 but it will use the neural cores or the GPUs for acceleration. So here I was expecting a similar, or maybe less, 3x performance.
So the CPU/GPU combination did only work at 3.5x, this is under impersonation though using Rosetta, but really interestingly, the DXO Deepprime was much better running on the neural cores than those 32 GPUs. DXO’s claim that its an “AI” based de-noise appears to mean that it runs better on the 16 neural cores.
Clearly software developers at both DXO and Adobe have much more capacity that they can squeeze out of this hardware, so its liable that these numbers will improve. But it also points to the importance of hardware acceleration in AI, and how custom chips can outperform ‘generic’ GPUs.
