Inspur OAM UBB Sets New Accelerator Platform Standard

Inspur unveiled a number of new technologies at the Open Compute Project (OCP) Regional Summit in Amsterdam this week. Perhaps the most impactful of these over the next 24 months is the Open Accelerator Module (OAM) Universal Baseboard (UBB.) We are going to use OAM UBB to keep that shorter. The OAM UBB is set to be the next-generation hyper-scale acceleration platform. Patrick, our Editor-in-Chief, has an interview and a more in-depth look at why that will go live early next week. In the meantime, he asked me to do a quick overview of the OAM UBB.

OCP OAI Background

At the OCP Summit 2019 earlier this year, OAM was a major theme. More broadly the Open Accelerator Infrastructure or OAI. Since its launch, STH has covered the OAM/ OAI a number of times:

• Intel Nervana NNP L-1000 OAM and System Topology a Threat to NVIDIA
• Facebook Zion Accelerator Platform for OAM
• Facebook OCP Accelerator Module OAM Launched
• Favored at Facebook Habana Labs Eyes AI Training and Inferencing
• Intel Nervana NNP-T for AI Training Hot Chips 31
• STH Interview with Bill Carter of OCP and John Hu of Inspur

The OAM Universal Baseboard is designed to help different organizations utilize the OAM ecosystem in different form factors. Open Accelerator Modules are most commonly GPUs or AI accelerators and the spec allows for different interconnect topologies.

Essentially, this is the “universal” platform for a next-generation NVIDIA DGX-1 that can use NVIDIA Tesla, but also accelerators from companies like Intel, AMD, and Habana labs supporting different scale-out protocols for each.

Over the past few quarters, this anti-lock-in platform that is seeking to pry the industry away from proprietary NVLink and NVSwitch architectures has gotten a lot of attention and support. Basically, everyone other than NVIDIA wanted this, and NVIDIA has come along begrudgingly.

As a result, we are expecting broad support over the next 12-36 months for OAI and OAM.

Inspur OAM UBB Contribution to the OCP Ecosystem

A key to the technology, beyond the modules, is a common motherboard platform that provides the connectivity for up to eight devices inside the system as well as external connectivity. At the OCP Regional Summit 2019, Inspur released its Universal Baseboard that allows for a number of accelerator topologies.

By providing this Universal Baseboard in a reference platform (the Inspur ON5388M5) there is now a relatively easy to source solution available in the OAI/ OAM market from a top 3 server vendor.

The OAM reference system includes not just the UBB, but also a full power and air cooling solution for modules.

With the UBB available, we are seeing a number of designs based on the solution. One of the first is the Baidu X-Man 4.0. This is an updated platform to the Baidu X-MAN Liquid Cooled 8-Way NVIDIA Tesla V100 Shelf that we saw previously using the Inspur UBB to support next-generation OAM form factor modules.

Final Words

STH will have more on this in an interview over the next few days. This is a major shift in the industry and the UBB will enable more rapid designs to come to market supporting OAM. Instead of highly proprietary NVLink solutions, there is now an open platform that supports a number of different accelerators.