1. Overview

The IPU-Machine: M2000 is a 1U compute platform for AI infrastructure and is scalable for both direct attach and switched systems up to a 64K-IPU scale-out configuration. The IPU-M2000 is characterised by the following high-level features:

  • 4x GC200 IPUs

    • ~1 petaFLOPS FP16.16 AI compute

    • 5,888 processor cores

    • 35,000 independent parallel threads

  • Up to ~260GB of memory comprised of:

    • Up to 256GB Streaming Memory™

    • 3.6GB In-Processor-Memory™

  • IPU-Fabric™ for compiled-in networking comprised of:

    • IPU-Link™ - 512Gbps for communication within IPU-PODs

    • GW-Link - 2x 100Gbps Gateway-Links for communication between IPU-PODs

    • Sync-Link - dedicated hardware signalling for BSP, low jitter on IPU to IPU synchronisation

    • Host-Link - PCIe Gen4 RoCEv2 NIC/SmartNIC Interface for IPU-M2000 to server communication