Sieve: Dynamic Expert-Aware PIM Acceleration for Evolving Mixture-of-Experts Models

TL;DR

This paper introduces Sieve, a runtime framework with a scheduler that dynamically optimizes PIM acceleration for evolving mixture-of-experts models, significantly improving throughput and interactivity.

Contribution

It presents a novel runtime framework and scheduler that adaptively partition expert execution between GPU and PIM based on runtime distributions, addressing load imbalance and communication issues.

Findings

Sieve improves throughput and interactivity by up to 1.6x over state-of-the-art PIM systems.

Modern MoE models exhibit bimodal token-to-expert distributions, impacting PIM efficiency.

The proposed scheduler effectively balances load and reduces communication overhead.

Abstract

Mixture-of-Experts (MoE) has become a dominant architecture for scaling large language models (LLMs). However, the execution characteristics of MoE inference are changing rapidly and increasingly mismatch the assumptions underlying existing Processing-in-Memory (PIM) systems. Prior PIM systems for LLMs rely on static rules to offload memory-bound operations to PIM, without accounting for the combined effects of load imbalance and inter-GPU communication. Meanwhile, modern MoE models activate fewer experts out of increasingly many, creating a bimodal expert distribution: a small set of experts receives many tokens, while a long tail of experts receives only one or a few. We identify a trend in modern MoE models toward increasingly bimodal token-to-expert distributions, quantify the resulting disparity in arithmetic intensity across experts, and show that this disparity dramatically reduces the efficiency of state-of-the-art PIM systems for LLMs. To address this problem, we propose a scheduler for serving MoE models on multi-GPU systems with attached HBM-PIM stacks. Our scheduler partitions expert execution between GPU and PIM based on runtime token-to-expert distributions, while jointly considering interconnect overhead, memory bandwidth, GPU throughput, and PIM throughput. Moreover, we propose Sieve, a runtime framework that employs the scheduler to coordinate execution across GPUs and their attached HBM-PIM stacks. Sieve overlaps GPU computation, PIM computation, and intra- and inter-device communication while preserving cross-device dependencies induced by expert parallelism. Sieve is evaluated on our cycle-accurate simulator based on Ramulator 2.0. Compared to state-of-the-art PIM systems for MoE, Sieve improves both throughput and interactivity by 1.3x, 1.3x, and 1.6x on Qwen3.5-397B-A17B, GPT-OSS-120B, and Qwen3-30B-A3B, respectively.