Back To Schedule
Wednesday, May 13 • 10:55am - 11:10am
Future Inference Accelerator Module Form Factor

Sign up or log in to save this to your schedule, view media, leave feedback and see who's attending!

To fulfill the needs of scaling FB infrastructure and improving performance, power, and efficiency, the development of specialized accelerators and holistic system-level solutions has shown great importance. In order to gain efficiency, thermal/mechanical design and management for those high density, small form factor modules is an important building block to ensure the overall efficiency of the solution. This talk will discuss mechanical and thermal considerations of both current and future generation of the platform for this particular use cases, provide opportunities for further efficiency gain and seek for future collaboration on standardizing and scaling the solution. The talk will also expand and share challenges of the current form factor (M.2 and dual M.2) and discuss further potential scalable form factors.

Topics that might be covered during this talk:
1. FB Glacier Point V2 architecture in YosemiteV2 platform
2. FB next generation architecture (code name TBD) in next generation platform
3. Comparison between the system generation and potential thermal efficiency and power efficiency gain
4. Challenges and Call out for Action:
Form factor impact from die size, heatsink size, orientation, new cooling technologies perspective
Potential intersection with PCIe Gen5
Package design vs. system design - how can we work together as a community

avatar for Yueming Li

Yueming Li

Thermal Engineer, Facebook

Michael Haken

Mechanical Engineer, Facebook
avatar for Hao Shen

Hao Shen

Hardware Engineer, facebook
Hao is hardware engineer at Facebook working on AI platform. His focus is on AI Hardware System, Accelerator enablement and large scale datacenter solution.

Wednesday May 13, 2020 10:55am - 11:10am
EW: Servers