Tasa: Thermal-aware 3D-Stacked Architecture Design with Bandwidth Sharing for LLM Inference

TL;DR

Tasa introduces a thermal-aware 3D-stacked architecture with bandwidth sharing for LLM inference, significantly improving scalability, thermal management, and performance over existing solutions.

Contribution

The paper presents Tasa, a heterogeneous 3D-stacked architecture with thermal optimization and bandwidth sharing, enhancing LLM inference efficiency and scalability.

Findings

Up to 5.55°C peak temperature reduction in 48-core configurations.

Achieved 2.85x and 2.21x speedup over GPU baselines for Llama-65B and GPT-3 66B.

Demonstrated improved thermal scalability and inference performance.

Abstract

The autoregressive decoding in LLMs is the major inference bottleneck due to the memory-intensive operations and limited hardware bandwidth. 3D-stacked architecture is a promising solution with significantly improved memory bandwidth, which vertically stacked multi DRAM dies on top of logic die. However, our experiments also show the 3D-stacked architecture faces severer thermal issues compared to 2D architecture, in terms of thermal temperature, gradient and scalability. To better exploit the potential of 3D-stacked architecture, we present Tasa, a heterogeneous architecture with cross-stack thermal optimizations to balance the temperature distribution and maximize the performance under the thermal constraints. High-performance core is designed for compute-intensive operations, while high-efficiency core is used for memory-intensive operators, e.g. attention layers. Furthermore, we propose a bandwidth sharing scheduling to improve the bandwidth utilization in such heterogeneous architecture. Extensive thermal experiments show that our Tasa architecture demonstrates greater scalability compared with the homogeneous 3D-stacked architecture, i.e. up to 5.55 $\tccentigrade$, 9.37 $\tccentigrade$, and 7.91 $\tccentigrade$ peak temperature reduction for 48, 60, and 72 core configurations. Our experimental for Llama-65B and GPT-3 66B inferences also demonstrate 2.85x and 2.21x speedup are obtained over the GPU baselines and state-of-the-art heterogeneous PIM-based LLM accelerator