Characterizing and Efficiently Accelerating Multimodal Generation Model Inference

Kaushik Ram Sadagopan, Yejin Lee, Basil Hosmer, Daniel HAZIZA, Carole-Jean Wu, Wen-tau Yih, Anna Sun, Bilge Acun, Changhan Wang, Charles David Hernandez, Christian Puhrsch, Driss Guessous, Francisco Massa, Jacob Kahn, Jeremy Francis Reizenstein, Jiaqi Zhai, Joe Isaacson, Joel Schlosser, Juan Pino, Leonid Shamis et al. (8 additional authors not shown)

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 multimodal 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 3.88×better baseline.