Reproduce “NITRO: High-Performance 3D NAND Flash-Based In-Storage Computing with Enhanced Activation Dataflow” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Sanghun Shin, Sangyeon Kim, Gisan Ji, Sungju Ryu
NITRO separates static model weights from temporary activations in LLM inference. It keeps weight-heavy linear layers in TLC NAND-PIM and uses DRAM for activation buffering and attention, with plane-aware mapping to parallelize NAND work. In simulation across decoder-only models, the optimized DRAM-PIM design reaches normalized latency of approximately 0.003 versus 3D-FPIM and 0.069 versus S-FLASH; these are not physical SSD measurements.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.