GEN-Graph: Heterogeneous PIM Accelerator for General Computational Patterns in Graph-based Dynamic Programming

TL;DR

GEN-Graph introduces a heterogeneous PIM chiplet with specialized tiles for matrix-centric and traversal-centric dynamic programming, significantly improving performance and energy efficiency in genomics and network analytics.

Contribution

It presents the first scalable, exact heterogeneous PIM architecture tailored to different DP computation patterns, combining recursive partitioning and reconfigurable logic.

Findings

Matrix tile achieves 42.8x speedup over NVIDIA H100 for APSP.

Traversal tile reaches 2.56 million reads/sec for short-reads.

Outperforms state-of-the-art accelerators by up to 2.56x in throughput.

Abstract

While graph-based dynamic programming (DP) is a cornerstone of genomics and network analytics, its efficiency is hampered by fundamentally conflicting computational patterns. Matrix-centric DP drives regular, compute-bound network analytics, while topology-centric DP handles irregular, memory-bound genomic traversals. These two categories of DP have substantially different computation patterns and dataflows, which makes it difficult for a single homogeneous processing-in-memory (PIM) architecture to efficiently support both. This work presents GEN-Graph, a novel heterogeneous PIM chiplet that integrates two types of specialized compute tiles within a 2.5D package: Matrix-tile, a processing-using-memory (PUM) tile optimized for matrix-centric workloads, such as all-pairs shortest path (APSP); and traversal-tile, a processing-near-memory (PNM) tile optimized for traversal-centric DP workloads, such as DNA sequence alignment. Our hardware-software co-design employs recursive partitioning and reconfigurable windowed bit-parallel logic to ensure exact computation. Results show the matrix tile achieves 42.8x speedup and 392x energy efficiency over the NVIDIA H100 GPU for APSP. For sequence-to-graph alignment, the traversal tile sustains 2.56 million reads/s (short-reads) and 39.3 thousand reads/s (long-reads), outperforming state-of-the-art accelerators by up to 2.56x in throughput. GEN-Graph provides the first scalable, exact solution for general DP dataflows by matching hardware specialization to algorithmic structure.