We brought significant architectural advancements and feature set improvements to A19/M5/M5 Pro and Max GPUs:
- Scalable GPU Neural Accelerators
- 2nd Generation Dynamic Caching Shader Architecture
- 3rd Generation Ray Tracing Acceleration
- The best GPU Driven Pipeline Architecture is even better
- Many new graphics features that makes games and Pro Apps even better
- Rate increases and new performance features throughout the design
We released two tech talks.
We released two tech talks today going over how to take advantage of the new architecture, features and associated developer tools.
Accelerate your machine learning workloads with the M5 and A19 GPUs
https://developer.apple.com/videos/play/tech-talks/111432/
https://www.youtube.com/watch?v=wgJX1HndGl0Boost
your graphics performance with the M5 and A19 GPUs
@gavkar here are the proper links:
Discover how to take advantage of the M5 and A19 GPUs to accelerate machine learning. Find out how to use the Neural Accelerators inside each GPU core to per...
@gavkar Ah, I’ll have to get an M5 to benchmark it then. I spent some time reverse engineering the format, will have to revisit that. I imagine smaller block sizes may help implement scattered writes.
BTW, what’s the use case? What applications/algorithms benefit from this feature?
@castano almost all games use compute shaders. When developers marked a texture “compute write” we disabled lossless compression. They could make it compressed again by “optimizeForGPU”.
On m5 we support scattered writes as well. Which means even for textures marked as “compute write” they get lossless compression.
Lossy compression (I am assuming you don’t mean ASTC, BC) is on the same HW path.
If I misunderstood you and you meant ASTC or BC then you should expect no changes.
@gavkar No, that's what I meant. I always hoped that tile shaders would prove useful to compress on tile eviction, even though in practice it did not perform well. But the compute use case is something I don't see how you could replicate with a custom codec.
I imagine you have to keep the block contents in local memory/registers and compress when evicted. Maybe the new dynamic scheduler is good enough to handle that sort of workload?
Gokhan Avkarogullari
Ignacio Castaño🍉