Zhantong Zhu

Kangbo Bai, Zhantong Zhu, Yifan Ding et al.

In large-scale distributed LLM training, communication between devices becomes the key performance bottleneck. Chiplet technology can integrate multiple dies into a package to scale-up node performance with higher bandwidth. Meanwhile, optical interconnect (OI) technology offers long-reach, high-bandwidth links, making it well suited for scale-out networks. The combination of these two technologies has the potential to overcome communication bottlenecks within and across packages. In this work, we present ChipLight, a cross-layer multi-objective design and optimization method for training clusters leveraging chiplet and OI. We first abstract an architecture model for such complex clusters, co-optimizing chiplet architecture, training parallel strategy, and OI network topology. Based on such models, we tailor the design space exploration flow by combining both black-box and white-box methodologies. Evaluated by our experimental results, ChipLight achieves significantly improved training efficiency and provides valuable design insights for the development of future training clusters.

Zhantong Zhu, Hongou Li, Wenjie Ren et al.

With the rapid advent of generative models, efficiently deploying these models on specialized hardware has become critical. Tensor Processing Units (TPUs) are designed to accelerate AI workloads, but their high power consumption necessitates innovations for improving efficiency. Compute-in-memory (CIM) has emerged as a promising paradigm with superior area and energy efficiency. In this work, we present a TPU architecture that integrates digital CIM to replace conventional digital systolic arrays in matrix multiply units (MXUs). We first establish a CIM-based TPU architecture model and simulator to evaluate the benefits of CIM for diverse generative model inference. Building upon the observed design insights, we further explore various CIM-based TPU architectural design choices. Up to 44.2% and 33.8% performance improvement for large language model and diffusion transformer inference, and 27.3x reduction in MXU energy consumption can be achieved with different design choices, compared to the baseline TPUv4i architecture.