Graphcore IPU

Realizes: machine intelligence workloads

Graphcore's Intelligence Processing Unit (IPU) is a massively parallel AI accelerator composed of thousands of SRAM-backed tile cores linked by an exchange-style interconnect, enabling sparse tensor graph processing workloads in IPU-POD16 and IPU-M2000 systems.

Examples

🔗

Transformer training on Graphcore IPU

Training a large transformer language model with fine-grained model parallelism and low-latency tensor math on Graphcore IPUs.

matrix multiplication; sparse and dense tensor contractions; collective inter-IPU communication high large ~1.5 pJ/MAC

🔗

Graphcore IPU-POD16 GNN workloads

The IPU-POD16 cluster powered by IPU-M2000 modules runs sparse tensor graph processing pipelines to accelerate GNN training and inference across large knowledge graphs with sub-microsecond token cycles.

Graph neural network message passing Sparse tensor graph traversal seconds large mJ