RoMe tests row-granularity memory for streaming LLM tensors

RoMe replaces cache-line transactions with row-granularity DRAM access for large contiguous LLM transfers, simplifying scheduling and reallocating pins to more channels.

Why it matters

RoMe is not flash, but it is a useful control point for HBF research on access granularity: it isolates the value of matching the memory interface to large tensor transfers without NAND latency or endurance effects.

Exact source title

RoMe: Row Granularity Access Memory System for Large Language Models

Authors

Hwayong Nam, Seungmin Baek, Jumin Kim, Michael Jaemin Kim, Jung Ho Ahn

RoMe tests row-granularity memory for streaming LLM tensors

RoMe is not flash, but it is a useful control point for HBF research on access granularity: it isolates the value of matching the memory interface to large tensor transfers without NAND latency or endurance effects.

Related records