Skip to main content

IPU (Intelligence Processing Unit) Architecture

What is an IPU

IPU (Intelligence Processing Unit) is Graphcore's (UK) intelligence processing unit, adopting a massively parallel MIMD (Multiple Instruction, Multiple Data) architecture. Unlike GPU's SIMD, each core independently executes different instructions.

Core features:

  • 1,472 independent processor cores (Bow IPU)
  • 900 MB on-chip SRAM (In-Processor Memory)
  • BSP (Bulk Synchronous Parallel) programming model
  • Ideal for fine-grained parallelism (e.g., recommendation systems, graph neural networks)

IPU Core Architecture

Bow IPU

  • 1,472 cores (Tile processors)
  • Each Tile = 1 core + 624KB SRAM
  • On-chip SRAM total 900 MB
  • Tile-to-Tile interconnect: Graphcore interconnect (IPU-Links)
  • Exchange chip (GW-Link) connects multiple IPUs

BSP Programming Model

  • Bulk Synchronous Parallel
  • Three-phase loop: compute + communicate + synchronize
  • Ideal for sparse graphs, dynamic graphs, fine-grained parallelism

Poplar SDK

  • C++ / Python SDK
  • TensorFlow / PyTorch integration
  • ONNX conversion tools

Mainstream IPU Products

IPUYearCoresSRAMCompute (FP16)TDP
Colossus MK1 GC220181,216304 MB-150W
Bow IPU GC20020201,472900 MB-150W
Bow Pod20204× GC2003.6 GB--
Bow Pod64202016× GC20014.4 GB--

IPU vs GPU

DimensionIPUGPU
ArchitectureMIMD (independent instructions)SIMT (unified instructions)
Parallel granularityFine (core-level independent)Coarse (warp-level)
MemoryOn-chip SRAM (900MB)HBM (80GB)
Programming modelBSPCUDA
Best forSparse graphs, dynamic graphs, recommendation systemsDense matrix (LLM, CNN)
Large model trainingMediumStrong
EcosystemSmallCUDA mature

IPU Use Cases

  • Recommendation systems (DLRM, embedding lookups)
  • Graph neural networks (GNN)
  • ✅ Molecular dynamics simulation
  • ✅ Financial modeling
  • ⚠️ LLM training (ecosystem and performance lag behind GPU)
  • ❌ Large-scale LLM inference (use GPU/TPU)

Commercial Deployments

  • Microsoft Azure (IPU instances)
  • Graphcore IPUs available in European and North American data centers
  • Glow compiler
  • Complementary to NVIDIA GPU (partial inference offload to IPU)

Detailed Product Pages