Экспериментально исследована архитектура и программное устройство графического процессора (ISP) в процессоре Apple M1 с помощью языка С и внешнего устройства обработки изображений. We validated the approach by comparing training results from simulated raw images against raw images captured in-lab. We found that processing images with an ISP improves accuracy by an average of 7.0% for a broad set of MobileNet CNNs. Our results indicate that the ISP has a more significant impact, and our results on the larger ResNet-50 and are consistent with this trend.