GR801

Status:

Development

GR801 is a radiation-hardened SoC designed to enable reliable artificial intelligence (AI) applications in space. The GR801 device is part of our GRAIN (Gaisler Research Artificial Intelligence NOEL-V) product line.

Overview

This page describes a running development and no guarantees can be given concerning future product availability. All information on this page is subject to change without notice. Please click this link to sign up to receive notifications about product and documentation updates.

The GR801 architecture consists of a neuromorphic processing engine (Brainchip AkidaTM 1.0), a single-core RISC-V RV64GC processor (NOEL-V) for system management, and several digital interfaces.

Performance target: >1TOPS below 0.5W

GR801 is a reliable computing platform for AI applications such as image recognition, autonomous navigation, and data analysis. The neuromorphic engine provides fast and energy-efficient inference, enabling fast on-board decisions.

Architecture

arrow down icon

Fault tolerance

arrow down icon

Key Tech Spec

arrow down icon

Target technology support

arrow down icon

Evaluation boards

arrow down icon

Ordering information

arrow down icon

Development Kit

arrow down icon

Licensing

arrow down icon

Software

arrow down icon

Tools

arrow down icon

Block diagram

arrow down icon

Related project

arrow down icon

Supported Hardware

arrow down icon

Configuration

arrow down icon

Reference Design

arrow down icon

Other resources

arrow down icon

Detailed features

arrow down icon

Package

  • Space grade thermally enhanced BGA,  23x23mm, 441 balls,  1mm pitch

NOEL-V RISC-V Processor

  • General-purpose RISC-V processor (RV64GC)
  • High-performance processing for traditional computing tasks
  • System management and configuration of the neuromorphic engine

AkidaTM neuromorphic processing accelerator

  • Akida provides acceleration of AI models using event-driven computing to minimize power consumption.  
  • Eight neural processing nodes connected in a mesh network
  • Each node consisting of four convolutional or fully connected engines
  • Each node supports 128 4x4 MACs, for total of 1024 MACs/clock
  • Hardware support for 1, 2, or 4-bit hybrid quantized weights
  • Multi-Pass Processing enables execution of large neural networks

Memory

  • 8 MB on-chip RAM
  • QSPI memory controller (2 chip selects)
  • 3.2 MB RAM private for the Akida unit

Interfaces

  • PCIe Gen 3 x4 port with capability to act both as a root port and as an endpoint
  • 1x 10/100/1000Mbit Ethernet interface (GMII and SGMII)
  • SpaceWire router - 4x external ports @ 200 Mbps
  • 4x CAN FD
  • 1x SPI (2 chip select)
  • 16x GPIO
  • 2x UARTs

Downloads

File

Category

Revision

Date

Access

No items found.

Frequently asked questions

No items found.