Hi,
In the full ranking results presented here: https://ai-benchmark.com/ranking_processors_detailed.html, the following values are provided for the A15 Bionic chipset for the SRGAN model:
- Int8: 485 ms
- Float16: 167 ms
I have a few questions regarding these data points.
1. As far as I know, int8 execution is not supported on any of the Apple chipsets, only float16 - how was this executed? On the CPU with float16/float32?
2. Are these runtime numbers including the input and output copy, i.e. are you using IOSurface types (https://developer.apple.com/documentation/iosurface) to avoid data copies between the CPU and NPU?
3. Why is the error 10 times larger (> 50) compared to any of the other chipsets?
In the full ranking results presented here: https://ai-benchmark.com/ranking_processors_detailed.html, the following values are provided for the A15 Bionic chipset for the SRGAN model:
- Int8: 485 ms
- Float16: 167 ms
I have a few questions regarding these data points.
1. As far as I know, int8 execution is not supported on any of the Apple chipsets, only float16 - how was this executed? On the CPU with float16/float32?
2. Are these runtime numbers including the input and output copy, i.e. are you using IOSurface types (https://developer.apple.com/documentation/iosurface) to avoid data copies between the CPU and NPU?
3. Why is the error 10 times larger (> 50) compared to any of the other chipsets?
