Characterizing and Efficiently Accelerating Multimodal Generation Model Inference

TL;DR

This paper analyzes the system design challenges of multimodal generative AI models, identifying bottlenecks like auto-regressive token generation and attention, and demonstrates optimization strategies that significantly improve inference efficiency.

Contribution

It characterizes multimodal generation models on real systems and proposes optimization techniques across applications, system software, and hardware to accelerate inference.

Findings

Auto-regressive token generation is a key latency bottleneck.

Memory-intensive attention and feed forward networks contribute to inference delays.

Optimization strategies can improve baseline performance by up to 3.88x.

Abstract

Generative artificial intelligence (AI) technology is revolutionizing the computing industry. Not only its applications have broadened to various sectors but also poses new system design and optimization opportunities. The technology is capable of understanding and responding in multiple modalities. However, the advanced capability currently comes with significant system resource demands. To sustainably scale generative AI capabilities to billions of users in the world, inference must be fast and efficient. This paper pinpoints key system design and optimization opportunities by characterizing a family of emerging multi-modal generation models on real systems. Auto-regressive token generation is a critical latency performance bottleneck, typically dominated by GPU idle time. In addition to memory-intensive attention across the generative AI models, linear operations constitute significant inference latency due to the feed forward networks in Transformer-based models. We demonstrate that state-of-the-art optimization levers, spanning from applications to system software and hardware, set a 3.88x better baseline.