Fusion Recurrent Neural Network (RNN) Accelerator

Overview

The NEUCHIPS RNNAccel-200/100 is a Deep Learning Accelerator IP which empower neural network inference for SoC/MCU/DSP. It is designed especially for ultra-low power applications and targeted toward mWatts edge devices. Including a neural network compression engine, it largely reduces memory footprint and memory access power. Supports popular AXI/AHB bus interfaces and companion tools to easy integrate, evaluate, and validate. RNNAccel is the smart solution to empower neural network on your edge processors.

Key Features

Multi-layer fusion neural network (each layer can be of type RNN, LSTM, GRU, or FC/MLP)
Multiple concurrently neural networks
Parameterized number of neurons for input/hidden/output layer
16-bit fixed-point input/output
8-bit/16-bit fixed-point weight/bias
NeuCompression: 5.3X-16X (6~2 bits per weight) smart parameter decoder
Configurable decimal point position for input, weight/bias and output of each layer
Configurable 16 16-bit x 16-bit or 32 16-bit x 8-bit MACs
Activation functions: Softsign, Sigmoid, ReLU, Tanh
AMBA3 AXI/AHB on-chip-bus compliant

Benefits

MAC utilization up to 99%
Energy efficiency 2.06 TOPS/W
Peak performance can scale up to 204.8 GOPS
High accuracy compression
Toolchain

Bit-accurate C simulation C
NeuCompression: smart parameter compression tool

Supporting Network Types

Vanilla Recurrent Neural Network (RNN)
Long Short-Term Memory (LSTM)
Fully Connected (FC)/ Multi-Layer Perceptron (MLP)
Gated Recurrent Unit (GRU)

Block Diagram

Applications

Recommendation system
eCommerce

Deliverables

RTL, Netlist

Technical Specifications

Request Info