Extreme-Low-Power Edge Inference Accelerator with ReRAM-based Compute

ChipFoundry BM Labs NVM Power-Up Design Contest Submission

Project Overview

This project implements a small matrix multiplication accelerator utilizing systolic arrays that, paired with BM Labs' ReRAM and Analog In-Memory Compute, enables extreme-low-power yet fast edge AI inference.

Key Innovation Points

• Leverage Analog In-Memory Compute: Utilize automatic quantization for low-power AI
• Small and Mighty: Slim design minimizes power consumption

Technical Highlights

• 8x8 byte matrix multiplication accelerator
• Wishbone bus interface for Caravel SoC integration
• Extreme-low-power operation

Architecture

Replicating Locally

Follow these steps to set up your environment and harden the design:

1. Clone the Repository:

git clone https://github.com/partcleda/caravel_user_neuromorphic_comp.git 

2. Prepare Your Environment:

cd caravel_user_neuromorphic_comp
make setup

3. Install IPM:

pip install cf-ipm

4. Install the Neuromorphic X1 IP:

ipm install Neuromorphic_X1_32x32

5. Edit Behavioral Model Name in IP:

The cocotb simulation flow uses verilog/includes/includes.rtl.caravel_user_project as its source files, which includes a path to the Neuromorphic IP behavioral model. In order to avoid making a second user_project_wrapper.v, it is simpler to modify the behavioral model module name from Neuromorphic_X1 to Neuromorphic_X1_wb to align with the stub that is used when actually hardening. With this change, the same user_project_wrapper.v works for both (cocotb) testbenching as well as hardening.

In other words, rename line 16...

File: ip/Neuromorphic_X1_32x32/hdl/beh_model/Neuromorphic_X1_Beh.v
16: module Neuromorphic_X1_wb (

6. Run Testbenches:

make cocotb-verify-all-rtl

7. Harden the Design:

make user_project_wrapper

Application: Extreme-Low-Power AI

This neuromorphic accelerator targets real-time inference in extremely power-limited environments:

• Deep space, sea inference where on-board power limited
• Low-cost chiplet integration with minimal heat and cost

Why This Design Wins

Innovation: Novel approach that exploits ReRAM's unique analog computing capabilities for in-memory compute, not just simple data storage.

Practicality: Hyper-focus on minimalist efficiency simplifes design and makes for low-risk integration in any project.

Differentiation: While most designs use NVM for basic memory, this leverages it for neural network inference acceleration.

Documentation

• Details about the Neuromorphic X1 IP: Neuromorphic X1 documentation
• Competition details: ChipFoundry BM Labs Challenge

License

This project is licensed under Apache 2.0 - see LICENSE file for details.