Profiling Apple Silicon Performance for ML Training

TL;DR

This paper evaluates Apple Silicon's performance in ML training, revealing significant differences from NVIDIA GPUs due to system-level factors and memory architecture, with implications for future ML hardware optimization.

Contribution

It provides a comprehensive analysis of Apple Silicon's ML training performance, highlighting the impact of Unified Memory and system-level factors compared to NVIDIA GPUs.

Findings

Apple Silicon shows a performance gap with NVIDIA GPUs in ML training.

System factors like page faults and power consumption significantly affect performance.

BLAS performance analysis explains part of the observed performance differences.

Abstract

Apple Silicon has attracted much attention for its performance and role in machine learning (ML) training. Unlike NVIDIA GPUs, which have traditionally dominated ML training, Apple Silicon has a significant difference in memory architecture. It uses Unified Memory, which integrates CPU and GPU memory instead of separate CPU memory and GPU VRAM. However, it is difficult to tell whether Unified Memory means more performance benefits. This paper investigates the performance differences by training several large language model (LLM) workloads end-to-end under different memory scenarios. The results show a significant performance gap between Apple Silicon and NVIDIA GPUs. This paper attributes this gap to system-level factors such as page faults, power consumption, and kernel launch time. In addition, the performance difference of basic linear algebra subprograms (BLAS) on the NVIDIA GPUs and Apple Silicon chips is analyzed to further explain the observed gap.