Expected MetalFX TAAU frame time?

Is there a reference for how much time should MetalFX temporal antialiasing + upscaling take? On a M1 Pro, for a test frame with an otherwise extremely short render time (see attached image, render takes ~600µs), the upscaler adds an additional ~10ms just for Metal FX temporal antialiasing + upscaling (see attached image). This obviously leaves very little time to keep the frame time under 16ms (60fps) and makes 8ms frame times (120fps) impossible.

The upscaling is working with an input render size of 1,904x1,370 and output size of 2,856x2,055, so it's not like the image is being upscaled to an outrageous size or anything.

Any thoughts? Is this to be expected? What can affect the Metal FX Temporal Upscaling frame time?

The way Apple designed MetalFX, seems to be hiding the latency of accessing the neural engine. I noticed that the GPU's clock speed is consistently at ~400 MHz with light rendering workloads; heavier workloads will access higher clock speeds and decrease MetalFX upscaling time proportionally. The key observation is that the Apple GPU makes clock speed as low as possible to finish your frame in time, and minimize power consumption. You can overlap a lot of your own computations in parallel to the MetalFX upscaling, or between gaps in commands that are issued by MetalFX.

Here's another way to look at it. My M1 Max GPU can theoretically consume 50 W. MetalFX is taking 4 ms when I'm rendering at 120 Hz (8.3 ms/frame). However, it is not consuming (4/8.33 x 50 =) 24 W. It is consuming 2 W. The total latency also doesn't change much with different resolutions; I mostly use 768x768 -> 1536x1536.

Expected MetalFX TAAU frame time?
 
 
Q