DASH gives HBF two concurrent paths into the GPU

DASH combines a direct GPU–HBF route with an HBF–HBM–GPU relay and uses both concurrently to deliver MoE expert weights, while managing immutable experts separately from mutable KV-cache traffic.

Why it matters

It moves beyond treating HBF as passive capacity and directly studies expert placement, early reads, and dual UCIe data paths. The reported 1.94× throughput and 1.90× end-to-end speedup are simulator results from a preprint, not silicon measurements.

Exact source title

Beyond Capacity: Scalable MoE LLM Inference via High-Bandwidth Flash with Direct GPU and HBM Paths

Authors

Seeyeon Kim, Juhyeong Jin, Joo-Young Kim

DASH gives HBF two concurrent paths into the GPU

It moves beyond treating HBF as passive capacity and directly studies expert placement, early reads, and dual UCIe data paths. The reported 1.94× throughput and 1.90× end-to-end speedup are simulator results from a preprint, not silicon measurements.

Related records