Best Practices for Metal on Apple Silicon: Architecture Migration, GPU Changes, and Optimization Techniques

This article discusses the impact of Apple Silicon on Metal applications and provides migration best practices for developers already familiar with Metal on iOS or macOS.

Architecture Migration

When Apple Silicon was introduced, Metal apps run natively on Intel but under Rosetta on Apple Silicon, incurring performance loss; rebuilding with the new macOS SDK eliminates Rosetta overhead and enables architecture‑specific optimizations.

Intel: code runs natively.

Apple Silicon: code runs on an optimized Rosetta layer.

Rebuilding with the new SDK removes Rosetta loss.

Apply Apple‑Silicon‑specific optimizations (see "Optimize Metal Performance for Apple Silicon Macs").

GPU Layer

Apple Silicon introduces new GPU capabilities and switches Metal from Immediate Mode Rendering ( IMR ) to Tile‑Based Deferred Rendering ( TBDR ), improving memory bandwidth and performance.

Key rendering modes:

IMR (Immediate Mode Rendering): simple but wasteful.

TBR (Tile Based Rendering): processes 32×32 tiles, still a form of deferred rendering.

TBDR (Tile Based Deferred Rendering): adds hidden‑surface removal ( HSR ) to render only visible pixels.

Benefits of TBDR include reduced memory bandwidth, register‑based blending, and elimination of unnecessary color/depth buffer reads.

OpenGL & OpenCL

Both frameworks are deprecated on macOS (OpenGL 4.1, OpenCL 1.1) and should be migrated to Metal.

Metal API Enhancements

Apple Silicon adds features such as Memoryless Render Targets, Programmable Blending, and On‑Chip MSAA Resolve, enabling forward‑deferred rendering, anti‑aliasing, and other performance optimizations.

Best Practices

Metal Feature Detection

Do not rely on static macros to differentiate iOS/macOS; query GPU capabilities at runtime to handle Apple Silicon correctly.

Load/Store Actions

Incorrectly setting loadAction or storeAction to dontCare can cause uninitialized memory or missing results; use load and appropriate store actions instead.

Coordinate Consistency

Multiple pipelines must compute vertex positions consistently; mismatched computePosition calls can break depth testing when using EQUAL .

Threadgroup Memory Synchronization

Use simdgroup_barrier or threadgroup_barrier to avoid race conditions in compute shaders; minimize barrier usage for performance.

Attachment Sampling

Sampling a depth or stencil attachment during the same render pass leads to read‑write hazards; create a copy of the attachment if sampling is required.

Rendering Consistency

Metal automatically upgrades loadAction set to DontCare to load , forces coordinate consistency, and snapshots depth attachments for apps built with older SDKs, though this incurs a performance cost.

Conclusion

Apple Silicon dramatically improves Metal performance and adds cross‑platform features, allowing developers to share code more easily across Apple devices. Further optimizations are detailed in the upcoming "Optimize Metal Performance for Apple Silicon Macs" session.