Tuesday, 10 October 2023

Data Centers At Meta: Heterogeneous Integration Driven By AI/ML And Network Applications

Last year, tech giants including Intel, Meta, Arm, Google Cloud, AMD, Qualcomm, TSMC and ASE formed the Chiplet consortium. A news article in Fierce Electronics said:

Several giant tech companies have joined hands to promote an open standard for chip components called chiplets and how they are crammed together in system-on-chip (SoC) designs deemed critical to a variety of future handheld and high-performance computers that power AI applications and much more.

The open standard, called Universal Chiplet Interconnect Express (UCIe), has been developed by Intel and clearly benefits Intel’s integrated device manufacturer (IDM) strategy as it builds new chip fabs in Arizona and Ohio and elsewhere outside the U.S.  Intel has been a prominent voice in the push to expand chip manufacturing outside of Taiwan and the rest of Asia where it is heavily focused today.

Intel has donated its UCIe standard to founding members in a new consortium that includes Intel along with Advanced Semiconductor Engineering, Taiwan Semiconductor Manufacturing Co., AMD, Arm, Google Cloud, Meta, Microsoft, Samsung and Qualcomm. The founders have already ratified UCIe 1.0 which covers the die-to-die physical layer, die-to-die protocols and software stacks which leverage the existing PCI Express (PCIe) and Compute Express Link (CXL) industry standards.

Ravi Agarwal, a technical sourcing manager at the Facebook/Meta Infrastructure group is responsible for driving advanced packaging architectures and foundry for both networking and AI/ML compute applications to meet Facebook’s future workloads. He is driving Chiplet Business Workstream in Open Domain-Specific Architecture (ODSA) Sub-Project within the Open Compute Project (OCP), working with ecosystem partners to enable a Chiplet marketplace. 

In a talk delivered for the IEEE Electronics Packaging Society (EPS) SFBA, he focused on heterogeneous integration for Artificial Intelligence, Machine Learning and network applications at Meta Infrastructure, and discussed implications for packaging and system-level considerations. In the talk he also shared some of the advanced packaging (chiplet) initiatives in which Meta is participating to develop an open ecosystem.

The talk is embedded below:

While the slides of this talk is not available, you can see slides of another talk he delivered here.

Related Posts