Debunking the CUDA Myth Towards GPU-based AI Systems

TL;DR

This paper evaluates Intel Gaudi NPUs as a potential alternative to NVIDIA GPUs in AI systems, highlighting competitive performance and energy efficiency, but noting software maturity challenges.

Contribution

It provides a comprehensive comparison of Gaudi NPUs and GPUs, including microbenchmarking and software optimization strategies, revealing their potential and current limitations.

Findings

Gaudi-2 achieves competitive performance with A100 in key AI workloads.

Gaudi NPU's energy efficiency is comparable to A100.

Software maturity remains a challenge for Gaudi NPUs.

Abstract

This paper presents a comprehensive evaluation of Intel Gaudi NPUs as an alternative to NVIDIA GPUs, which is currently the de facto standard in AI system design. First, we create a suite of microbenchmarks to compare Intel Gaudi-2 with NVIDIA A100, showing that Gaudi-2 achieves competitive performance not only in primitive AI compute, memory, and communication operations but also in executing several important AI workloads end-to-end. We then assess Gaudi NPU's programmability by discussing several software-level optimization strategies to employ for implementing critical FBGEMM operators and vLLM, evaluating their efficiency against GPU-optimized counterparts. Results indicate that Gaudi-2 achieves energy efficiency comparable to A100, though there are notable areas for improvement in terms of software maturity. Overall, we conclude that, with effective integration into high-level AI frameworks, Gaudi NPUs could challenge NVIDIA GPU's dominance in the AI server market, though further improvements are necessary to fully compete with NVIDIA's robust software ecosystem.