Skip to main content

Graphcore IPU

Vendor: Graphcore

Category: Innovative Architecture / Machine Intelligence

Architecture: IPU (Intelligence Processing Unit)

Introduction

A UK AI chip pioneer. The IPU uses a massively parallel architecture, optimized specifically for sparse computation and graph neural networks, delivering outstanding performance on specific AI workloads.

Specifications

ModelComputeMemoryInterfaceTDPProcess
Bow IPU350 TFLOPS (FP16)900MB on-chip SRAMIPU-Fabric Interconnect350W7nm
Mk2 IPU250 TFLOPS (FP16)900MB on-chip SRAMIPU-Fabric Interconnect275W7nm

Official Website

Visit Official Website

Driver Downloads

Linux

OS Support

WindowsLinuxmacOSAndroid

Version History

VersionRelease DateDescription
PopTorch 2.02024PyTorch 2.x compatibility

Performance Benchmarks

ModelTaskPerformance Metric
Bow IPUBERT-Large Training~85% H100 efficiency
Bow IPUPopART Graph TrainingSupports 1B+ parameter models
Mk2 IPUGeneral AIPrevious generation product

Pricing Information

ModelReference PriceNotes
Bow IPUContact salesSold by RapidAI (China subsidiary)
Mk2 IPUContact salesGradually replaced by Bow

Quick Setup

Linux

# 1. Install Poplar SDK (Graphcore programming framework)
pip install poplar
pip install popart

# 2. Verify
gc-info

Code Examples

Python (PopART)

import popart
import popart.ir as pir
import torch

# Use PopART IPU backend
ir = pir.Ir()
with ir.main_graph:
x = pir.ops.host_store(pir.constant([1.0, 2.0]))

Architecture Highlights

  • IPU (Intelligence Processing Unit): A processor designed for large model training, with intensive on-chip communication
  • Bow Architecture: Second-generation IPU, introducing memory optimizations and computational efficiency improvements
  • Poplar + PopART: Graphcore's unified compiler and runtime framework

Model Compatibility

Model/FrameworkSupport StatusNotes
PopART (PyTorch)✅ NativeBest support
TensorFlowPopART backend
PaddlePaddle⚠️Adaptation in progress
Large Model TrainingDistributed training support

If you are evaluating alternatives, the following products may also fit your scenario: