Uploaded June 2026 | Updated September 2026, 1 week ago

---

Unlock powerful machine learning performance with the Metal Tensor API and Metal Performance Primitives (MPP) Tensor Ops library. Discover how to create portable operations that take advantage of Neural Accelerators in Apple M5 and A19 GPUs. Learn to build custom machine learning kernels for your Core AI applications, and find out how to work effectively with quantized data formats and GPU memory optimization.

Explore related documentation, sample code, and more:
Machine learning passes: developer.apple.com/documentation/Metal/machine-learning-passes
Metal Performance Shaders: developer.apple.com/documentation/MetalPerformanceShaders
Download the Metal Performance Primitives (MPP) Programming Guide: developer.apple.com/download/files/Metal-Performance-Primitives-Programming-Guide.pdf
Running inline ML operations in a shader with Metal 4: developer.apple.com/documentation/Metal/running-inline-ml-operations-in-a-shader-with-metal-4
Accelerate your machine learning workloads with the M5 and A19 GPUs: developer.apple.com/videos/play/tech-talks/111432
Combine Metal 4 machine learning and graphics: developer.apple.com/videos/play/wwdc2025/262

00:00 - Introduction
00:21 - Apple's ML software stack
02:25 - Managing quantized data
04:23 - Multi-plane tensors
05:17 - Quantized matrix multiplication
09:31 - Building advanced ops
13:35 - Integrating custom ops into Core AI
15:25 - Next steps

More Apple Developer resources:
Video sessions: apple.co/VideoSessions
Documentation: apple.co/DeveloperDocs
Forums: apple.co/DeveloperForums
App: apple.co/DeveloperApp