01
Apple ML software stack:
Focuses on Core AI, MLX, Metal.
Graphics & Games / 16:13
Optimize custom machine learning operations with Metal tensors
Graphics & GamesAI & Machine Learningiosipadosmacostvosvisionoswatchos
Key Points
- Covers powerful machine learning performance with the Metal Tensor API and Metal Performance Primitives (MPP) Tensor Ops library.
- Shows how to create portable operations that take advantage of Neural Accelerators in Apple M5 and A19 GPUs.
- Explains how to build custom machine learning kernels for the Core AI applications, and find out how to work effectively with quantized data formats and GPU.
- The session moves through Apple ML software stack, Managing quantized data, Multi-plane tensors, Quantized matrix multiplication, Building advanced ops, Integrating custom ops into Core AI.
- Key concepts include Metal, Metal tensors, GPUs, Core AI, MLX, TensorOps, iOS 26, iOS 27.
- Platform coverage: ios, ipados, macos, tvos, visionos, watchos.
Condensed Flow
02
Managing quantized data:
Focuses on iOS 26, iOS 27, MTLTensor.
03
Quantized matrix multiplication:
Focuses on Metal, Metal tensors, macOS 27.
04
Building advanced ops:
Focuses on TensorOps, FlashAttention.
05
Integrating custom ops into Core AI:
Focuses on Core AI, Metal, PyTorch.
More Detail
Additional details
- Detailed flow: Apple ML software stack -> Managing quantized data -> Multi-plane tensors -> Quantized matrix multiplication -> Building advanced ops -> Integrating custom ops into Core AI.
- APIs and concepts to recognize: Metal, Metal tensors, GPUs, Core AI, MLX, TensorOps, iOS 26, iOS 27, MTLTensor, macOS 27, FlashAttention, PyTorch.
- Version and support notes focus on TensorOps, iOS 27, MTLTensor.
- Implementation focus: Core AI, MLX, Metal, PyTorch, TensorOps, FlashAttention.