compute-in-memory 开源项目推荐（25 个）

BUAA-CI-LAB /Literatures-on-SRAM-based-CIM

基于SRAM的内存计算（CIM）研究阅读清单。

EzraWolf /TinyMOA-IHP26a

IHP26a TinyTapeout 实现的 RISC-V CPU，集成基于 SRAM 的存内计算（CIM）加速器，用于执行高效模拟矩阵乘法。

EzraWolf /SoC

LLM 推理 SoC

Scophield /snn-structural-evolution

从 ANN 到事件驱动 SNN 的四阶段结构演化极简 PyTorch 示例：阶段 0（基线 ANN）→ 阶段 1（二值化）→ 阶段 2（时序扩展）→ 阶段 3（时序累积）→ 阶段 4（复位与稀疏性控制）。

TimurIbrayev /DUB_Sparsity_for_XBars

用于交叉阵列的 DUB 稀疏性

Joschua-Conrad /PSumSim

PSumSim：用于模拟矩阵-向量乘法器的部分和量化模拟器

IamSid44 /cim-llm-accelerator

Hybrid LLM accelerator: systolic array for prefill, 8T-SRAM in-memory core for decode

shuijingzhicheng /xinglang

从零诞生的存算一体编程语言与基座 ⚛ 一切语言的源头也是尽头。10语言源头汇合(C/C#/Rust/Python/Lisp/Haskell/Erlang/APL/Q#/J)、统一XL01字节码、自研编译器栈、自举、从零训练10M参数线性注意力基座。The origin and the end of all languages. Storage-compute unity, born from scratch.

jakkarin-promsee /bio-analog-cpu

Two brains on one analog substrate from ~80% unsupervised SCFF bulk + ~20% closed-form SLDA namer: the math model for a forward-only, on-chip continual learner. Behavioral simulation, no silicon. Draft 6.0 = the "baby neocortex," validated across 11 phases.

cpldcpu /DRAM_exploration

Some experiments to perform parallel data operations (compute-in-memory) on a 1980ies DRAM chip controlled with a CH32V003 RISC-V MCU

HyperMR-project /HyperMR

HyperMR: Efficient Hypergraph-enhanced Matrix Storage on Compute-in-Memory Architecture. This work is presented at SIGMOD 2025.

ceedot-rock /pmcp

PmCp — Persistent Memory Core Processing: compute-in-persistent-memory runtime with a deterministic extent ISA, crash-safe journal, and Chamber-signed provenance

RicardoLi666 /RicardoLi666

Witmem Technology / 知存科技 employee referral, careers and recruiting information

williamhwangweiju /projection-sensitivity-mapping

Sensitivity-aware placement of GPT-2 onto aging analog in-memory computing (AIMC) tiles: hardware-aware fine-tuning, noise-sensitivity profiling, tile-aging simulation, and lifetime evaluation with IBM AIHWKit.

Mourya-wizard /tile-aware-tensor-train-compute-in-memory

Tile-budgeted Tensor-Train rank allocation for PIM / compute-in-memory crossbars: 203 tiles vs 254 (param budget) and 601 (uniform eps) to stay within 1 pt of dense ResNet-18 on CIFAR-10 (91.49%, 698 tiles at T=128). Also DS-CNN keyword spotting: 62 vs 70 dense tiles. TT-SVD, tile cost model, pareto sweeps.

ong-labs /tessellate

Closed-loop HW/SW co-design for compute-in-memory accelerators — NSGA-II searches per-layer weight precision and column pruning, verified by real PyTorch QAT and physics simulation. The LLM explains the search; it never produces the numbers.

buicongnguyen /SystemDesign

Visual system design interview atlas for backend, systems, hardware, embedded, and ACiM NPU design

KafCoppelia /witmem-mlir-cim

A simple CIM compiler based on MLIR

Lewis-panda /pim-foundry

Research framework for CIM/PIM tensor execution, heterogeneous simulation, and reproducible comparisons.

api-evangelist /houmo

houmo — independent third-party profile of a public API surface, by API Evangelist. HOUMO.AI (后摩智能) is a Chinese fabless semiconductor company founded in 2020 that develops compute-in-memory (CIM / storage-and-computation integrated) edge AI chips for running large language models and other AI workloads on endpoint devices.