ANEForge: Python for direct computation on the Apple Neural Engine
A new Python package called ANEForge allows developers to directly program the Apple Neural Engine (ANE) without relying on CoreML. This bypass enables more efficient use of the ANE, which is the dedicated neural accelerator in Apple devices. ANEForge compiles tensor graphs into ANE programs, supporting operations like fused attention, various weight formats, and even training steps directly on the engine. This allows for significantly faster execution of models, such as a ResNet-18 forward pass completing in 0.33ms. AI