The Supermicro Ultra server expands the Graphcore IPU-POD configuration

     Following the successful certification of the first Supermicro Ultra system, Graphcore customers can now choose Supermicro servers as part of their IPU-POD configuration through Graphcore Elite Partners. The AS-1124US-TNRP is equipped with the latest third-generation AMD EPYC™ processor and is the first Supermicro's highest-performing enterprise-class Ultra server approved for use in IPU-POD systems.

     As part of the Graphcore IPU-POD, Supermicro Ultra servers will help innovators push the boundaries of machine intelligence to develop and deploy state-of-the-art models that accelerate today's most widely used AI applications.

     Tom Wilson, vice president of Product Management at Graphcore, said: "Supermicro is one of the most trusted companies in the industry, and its high-performance Ultra servers are the perfect complement to Graphcore's smart processors and scale-out systems designed specifically for AI. Graphcore continues to expand its offerings and provide customers with more choice, and the addition of the IPU-POD option to Supermicro servers is another milestone in this process and a testament to our focus on creating high-performance, flexible and reliable systems."

     The IPU-M2000 and the host server in the IPU-POD system can be configured at different rates to help optimize the TCO (total cost of ownership) depending on the different demands placed on the server by a particular machine intelligence workload. Server-intensive applications such as computer vision, for example, can benefit from a higher server-IPU ratio than natural language processing workloads.

     Raju Penumatcha, senior vice president and chief product officer at Supermicro, said: "Graphcore has fully considered the architecture of its IPU-POD data center system and how to maximize the use of different servers for different AI workloads. We expect that Supermicro's Ultra server and other servers, when used in conjunction with features such as variable IPU-server ratios, will yield amazing results."

     IPU-POD is Graphcore's scale-out machine intelligence solution based on multiple instances of IPU-M2000 and a range of approved host servers and switches. The IPU-M2000 is a 1U data center AI blade capable of delivering 1 PetaFlop of AI computing.

     IPU-Fabric provides high bandwidth connectivity and manages compiled communications and computation through the Poplar software platform. Graphcore currently offers IPU-Pods as POD4, POD16, and POD64.

     The IPU-POD16 DA (Direct Attach) has four IPU-M2000 directly connected to the host server, making it an ideal choice for AI engineers to start IPU evaluation, proof-of-concept development, and piloting. With 16 IPU-M2000, 1 to 4 host servers and 2 switches, the IPU-POD64 is ideal for scaling out larger models and production workloads.