Search papers, labs, and topics across Lattice.
This paper introduces FLINT, a novel substrate that leverages high-bandwidth flash (HBF) technology to enhance the capacity of large language model (LLM) inference by addressing critical limitations in memory capacity and management. By implementing a dynamic burst-buffer controller, a phantom-plane refresh mechanism, and a read-only flash translation layer, FLINT optimizes the interaction between HBF and accelerator systems, effectively mitigating latency and maximizing throughput. The key result demonstrates that FLINT significantly improves inference performance while enabling the deployment of larger models in constrained environments.
FLINT transforms LLM inference by integrating high-bandwidth flash, overcoming memory constraints that limit model deployment and performance.
LLM inference is increasingly constrained by accelerator memory capacity rather than compute throughput. This constraint is especially acute in single-accelerator and small-node inference systems, where limited on-package memory capacity restricts the size of deployable models. HBF is an emerging 3D-stacked NAND flash technology that provides multi-terabyte near-accelerator capacity, making it a promising capacity tier for storing LLM weights. However, existing HBF-based proposals face three adoption challenges: they (1) rely on coarse-grained static prefetching for LLM weights aiming to hide the microsecond-level read latency of the NAND flash device while maximizing HBF's read throughput, (2) expose NAND flash management tasks (e.g., refresh operations) to the accelerator-visible critical inference path, and (3) miss optimization opportunities to specialize and optimize the flash-management mechanisms to the workload behavior. Our goal is to design an efficient HBF substrate that integrates HBF as a memory-capacity tier alongside HBM while addressing these three challenges. To this end, we propose FLINT, a workload-driven HBF substrate for capacity-scalable LLM inference. FLINT introduces three mechanisms: (1) a hardware burst-buffer controller that dynamically coalesces and pipelines HBF reads aiming to utilize existing NAND flash buffers while sustaining high HBF bandwidth, (2) a phantom-plane refresh mechanism, which removes refresh from the critical inference path by moving refresh-related NAND flash operations outside the read foreground back via low-cost resource duplication, and (3) a read-only FTL, which replaces SSD-class support for arbitrary writes with a compact table that translates logical weight bursts to physical HBF locations.
