MetalFX System Requirements?

I just tried running MetalFX upscaling on my A15 device and isn't supported (spatial or temporal). However, M1 should work even though it's a lower device family. This is not good because an iPhone has a weaker GPU, and can't render to larger resolutions natively with software ray tracing. MetalFX upscaling would help the most on smaller chips.

Hopefully Apple will enable it on iPhones at some point, once they have a better GPU. Could someone who owns an iPhone 14 Pro tell whether the A16 supports MFX? Take this with a grain of salt, but my educated guess is, Apple will support MFX on iPhone with the A17/M3 chip series. They will probably have hardware-accelerated ray tracing from Imagination Technologies, and upscaling is especially needed for ray traced contexts. For example, Nvidia launched DLSS to complement RTX.

As an alternative, you could technically port AMD FSR to iOS, and that's probably what developers will do anyway. I don't recommend this (not legal), but someone could partially reverse-engineer MetalFX just to run it on A14/A15. It might use a neural network (like XeSS), using simdgroup_matrix with half precision. But it would be much easier to run the MetalFX that's optimized for M1 (instead of FSR), and do so legally.

To reiterate, I DO NOT encourage that anybody try to reverse-engineer closed source code. I was just stating that as an example, in a theoretical sense.

However, in a theoretical sense (again), we can compare the processing power of a processor that does and doesn't support MetalFX.

• A15 with 5 cores is 1.5 TFLOPS F32, or 3.0 TFLOPS F16 (my guess to what MFX is bottlenecked by).
• M1 with 7 cores is 2.2 TFLOPS F32, or 4.4 TFLOPS F16.
• A16 could be around 1.9 TFLOPS F32, or 3.8 TFLOPS F16 (based on a recent AnTuTu benchmark). That's SO close to M1, so Apple might enable it. If not, the next generation (A17) will surpass 7-core M1.