The Building Block for Machine Intelligence Scale-Out
IPU-PODTM systems are designed to accelerate large and demanding machine learning models for flexible and efficient scale out.
The IPU-POD64 is a single rack configuration featuring 16 IPU-M2000TM compute blades, based on the innovative GC200 Intelligence Processing Unit (IPU). The IPU-POD64 can deliver up to 16 petaFLOPS of AI compute.
The whole system, hardware and software, has been architected together. IPU-POD64 supports standard frameworks and protocols to enable smooth integration into existing data center environments. Innovators can focus on deploying their AI workloads at scale, using familiar tools while benefitting from cutting-edge performance.
Disaggregation for customised compute
Machine intelligence workloads have very different compute demands. For production deployment, optimizing the ratio of AI to host compute can help to maximize performance, while improving total cost of ownership. IPU-POD systems allow flexible mapping of the number of servers and switches to the requisite number of IPU-M2000 platforms, so deployment is better tailored to production AI workloads. IPU-POD64 supports configurations of 1-4 servers.
Communication architecture built for scaling
Efficient data access and transfer can unlock greater AI performance. IPU-FabricTM is an innovative communication architecture for system-wide data transfer, extending high-speed interconnect within individual IPUs, across IPU-M2000s, between IPU-PODs and throughout the data center. IPU-Fabric delivers high-performance low-latency communication to maximize AI application efficiency and is built to work with standard data center communication technologies.
Built for AI developers
IPU-POD systems support industry-standard software tools. Developers can work with frameworks such as TensorFlow, PyTorch, PyTorch Lightning and Keras, as well as open standards like ONNX and model libraries like Hugging Face.
For deeper control and maximum performance, the Poplar framework enables direct IPU programming in Python and C++. Poplar allows effortless scaling of models across many IPUs without adding development complexity, so developers can focus on the accuracy and performance of their application.
At Graphcore we put power in the hands of AI developers allowing them to innovate. Our software stack supports industry open standards and is open source.
Access to AI expertise
Graphcore has a global network of partners to assist users of IPU-PODs all the way from installation and application development through to production deployment. For documentation and other help, visit our website.
Ease of deployment has been a paramount consideration in designing the IPU-POD. The result is a solution that supports standard hardware and software interfaces and protocols, and integrates effectively with existing data center infrastructures.
IPU-PODs support a rich suite of software and tools for management and visualization based on industry-standard open source software and open APIs including OpenBMC, Redfish DTMF, IPMI over LAN, Prometheus, and Grafana.
Industry-proven management tools
Docker and Kubernetes support makes it simple to automate application deployment, scaling, and management of IPU-PODs. Virtual-IPUTM technology offers secure provisioning of IPUs to different tenants and workloads. Developers can build model replicas within and across multiple IPU-PODs and provision IPUs across many IPU- PODs for very large models.
IPU-PODs have an easy-to-use, intuitive web GUI for simplified IPU resource management. Engineers can manage status, perform system tests, and provision IPUs for workloads. IPU-PODs also integrate with a variety of cloud provisioning and management stacks, including VMWare’s Radium.
Anda butuh bantuan mengenai produk Graphcore ? Konsultasikan segera kepada Sistech Kharisma
Mohon isi form berikut ini. Team kami akan segera menghubungi Anda