Specifications

Overview

The Graphcore C600 is a PCIe Gen4 dual-slot card built around the Bow (Mk2) IPU. Where a GPU schedules warps of threads onto SIMT lanes, the IPU is a MIMD machine: 1,472 independent IPU-Cores, each running its own program, giving 8,832 concurrently executing threads. That model is unusually well suited to irregular, sparse and graph-structured computation — GNNs, sparse attention, recommendation models and scientific simulation — where SIMT hardware idles on divergence.

Each IPU-Core is paired with its own slice of In-Processor Memory, giving the die 900 MB of on-chip SRAM instead of a conventional cache hierarchy. Placing the whole working set in SRAM removes the unpredictable latency of a cache miss, which is why the platform is often chosen for latency-sensitive and deterministic workloads. The card draws 185 W and scales out over IPU-Link into Bow-POD64 rack systems.

Graphcore's Poplar SDK provides the compiler, and PopTorch / TensorFlow bindings let standard framework models be targeted at the hardware without a full rewrite. QS Compute can source C600 cards and Bow-POD systems for research and enterprise buyers — contact us for availability and pricing.