For more performance results, visit our Performance Results page
Start growing your AI compute capacity in the cloud or in your datacenter with IPU-POD128.
IPU-POD128 is designed for straightforward deployment, integrating effectively with standard datacenter infrastructure, including VMWare virtualization, OpenStack. Slurm and Kubernetes support, so it's simple to automate application deployment, scaling, and management. Virtual-IPU™ technology offers secure multi-tenancy. Developers can build model replicas within and across multiple IPU-PODs and provision IPUs across many IPU-PODs for very large models.
World-class results whether you want to explore innovative models and new possibilities, faster time to train, higher throughput or performance per TCO dollar.
IPUs | 128x GC200 IPUs |
IPU-M2000s | 32x IPU-M2000s |
Memory | 115.2GB In-Processor-Memory™ and up to 8.2TB Streaming Memory |
Performance | 32 petaFLOPS FP16.16 8 petaFLOPS FP32 |
IPU Cores | 188,416 |
Threads | 1,130,496 |
IPU-Fabric | 2.8Tbps |
Host-Link | 100 GE RoCEv2 |
Software |
Poplar TensorFlow, PyTorch, PyTorch Lightning, Keras, Paddle Paddle, Hugging Face, ONNX, HALO OpenBMC, Redfish DTMF, IPMI over LAN, Prometheus, and Grafana Slurm, Kubernetes OpenStack, VMware ESG |
System Weight | 900kg + Host servers and switches |
System Dimensions | 32U + Host servers and switches |
Host Server | Selection of approved host servers from Graphcore partners |
Thermal | Air-Cooled |
For more performance results, visit our Performance Results page