MLX, Core ML and Torch Performance on Apple Silicon

MLX vs PyTorch Performance on Apple Silicon for Training Core ML vs PyTorch Performance on Apple Silicon for Inference

Introduction

Comparing MLX and PyTorch on Apple Silicon

MLX uses Metal Performance Shaders (MPS) and PyTorch uses Metal Performance Shaders (MPS) and Metal Performance Shaders Graph (MPSGraph) on Apple Silicon.

PyTorch only uses CPU on Apple Silicon.

Comparing Core ML vs PyTorch on Apple Silicon

Core ML models uses the Neural Engine as well as the GPU for computation.

Environment

Python 3.10.15 on conda-forge

Clang 17.0.6 on darwin

Libraries

• PyTorch: torch 2.4.0
• TorchVision: torchvision 0.19.0
• Transformers: transformers 4.46.2
• MLX: mlx 0.20.0
• MLX-data: mlx-data 0.0.2
• coremltools: 8.1
• NumPy: numpy 2.0.2

Results Training

device	processor	ram	model	batch_size	epochs	mlx (s)	torch (s)
MBA	M1	8 GB	resnet-18	32	2	108.29	1176.04
MBP	M1	16 GB	resnet-18	32	2
MBA	M2	8 GB	resnet-18	32	2	59.11	1036.72

Results Inference

device	processor	ram	images	model	coreML	torch
MBA	M1	8 GB	2023	resnet-152	8.15	119.5
MBP	M1	16 GB	2023	resnet-152	NIL	NIL
MBA	M2	8 GB	2023	resnet-152	4.88	70.45