<img height="1" width="1" style="display:none" src="https://www.facebook.com/tr?id=145304570664993&amp;ev=PageView&amp;noscript=1">
Supermicro Ultra servers now available in Graphcore IPU-POD_logos

Mar 25, 2021

Supermicro Ultra servers now available in Graphcore IPU-POD

Written By:

Victoria Rege

Join the IPU conversation

Join our Graphcore community for free. Get help and share knowledge, find tutorials and tools that will help you grow.

Join on Slack

Putting Graphcore systems into the hands of AI innovators is about more than just the technology that we make ourselves – the IPU-M2000 and Poplar software – we deliver a portfolio of datacenter IPU-POD products with a complete reference architecture to help customers build AI compute systems at scale.

Our technology ecosystem includes many of the biggest, most trusted names in the industry, giving customers choice about how they buy, build and operate every part of their Graphcore system.

We are delighted to announce that Supermicro Ultra servers are now available from Graphcore Elite Partners as part of the IPU-POD product range.

The first server to be qualified from Supermicro’s highest performance enterprise-class Ultra range is the AS-1124US-TNRP. 

The Ultra is the ideal companion to Graphcore’s IPU compute systems, both for today’s most demanding AI applications, and for developing tomorrow’s state-of-the-art models.

In each IPU-POD system, servers work alongside our powerful IPU-M2000s, hosting the Poplar software, converting the chosen job into a graph and distributing it across IPUs, managing the flow of data from the network and storage, and performing management functions including allocation of IPUs via Virtual-IPU, as well as general hardware management and monitoring.

In addition to giving users the widest range of hardware options, we are committed to delivering the flexibility that lets them optimise their setup around specific workload requirements.

Within an IPU-POD setup - such as the IPU-POD64 - Supermicro Ultra servers and IPU-M2000 Machines can be configured in different ratios, depending on the most commonly run workloads. Computer vision, for example, tends to be a server-intense application so customers may wish to specify a higher number of Ultras. Natural language processing may require a lower server-to-IPU ratio.

This flexibility is possible thanks to Graphcore’s disaggregated server design. Rather than bundling the server and AI compute system together in the same box in a prescribed ratio, the two are separate - giving customers the option of how to specify and configure their setup.

Server disaggregation brings a number of other benefits, including the ability to physically optimise the arrangement of IPU-M2000s, servers and switches that make up a typical IPU-POD to maximise rack power utilisation. This is further aided by the slim, 1U form factor of the IPU-M2000 and 1U and 2U options in the Supermicro Ultra range.

The IPU-POD portfolio includes IPU-POD4 Direct Attach (DA), IPU-POD16  Direct Attach (DA) where one or four IPU-M2000s respectively are directly connected to a host server, and IPU-POD64 with 16 IPU-M2000s, 1-4 host servers and two switches.

The newly qualified Ultra servers join Graphcore’s list of approved servers, alongside Dell’s PowerEdge R6525. Further servers from Supermicro and other manufacturers will be added to the list as they are qualified.

Supermicro servers can be specified as part of a Graphcore IPU-POD through the following vendors: 2CRSi, Boston, Digital China, Meadowgate Technologies, Megware, and Wildflower.