Mr. Terrence TAN
Senior Director, Product and Test Engineering, Microsoft Silicon Engineering Penang, Microsoft
Terrence is currently the Senior Director of Product and Test Engineering at Microsoft SMPE Organization (Silicon, Manufacturing and Packaging) stationed in Penang. In his current role at Microsoft SMPE Penang (which started 1.5years ago), Terrence is responsible for NPI and Production Ramp for Microsoft Custom Server and AI chip, which was announced by the CEO, Satya in Microsoft Ignite September 2023. In addition, Terrence it also the chiplet interconnect test/debug architect at Microsoft SMPE organization.
Prior to Microsoft, Terrence spent 17years at Intel Penang in similar areas and was deeply involved in development of Intel's 3D stack manufacturing, test and debug capabilities.
Presentation Title
Leveraging AI/ML and Telemetry Capabilities at NPI and Production Phase to Optimize Silicon Quality, Power/Performance and Cost at Every Layer of Stacks
Abstract
Silicon is the workhorse of AI and Cloud computing today. Microsoft Ignite in September 2023 unveiled two custom designed chips: the Microsoft Azure Maia AI Accelerator, optimized for AI tasks and the Microsoft Azure Cobalt CPU, an ARM based processor for general purpose compute workload on Microsoft Cloud. With this, Microsoft can now deliver a full vertical stack system from silicon, software and server to racks and cooling systems in order to meet the AI demand. Silicon Quality/Reliability, Power/Performance, Cost/ TCO (Total Cost of Ownership) are 3 major KPIs we can customize and optimize at every layer of the vertical stacks - starting from chip fabrication at foundry, silicon test/assembly/packaging/qual at OSATs, building and testing custom server boards at System Integrator, placing the server boards with tailor made racks to be deployed to Microsoft datacenters.
This presentation will walk through how AI/ML capabilities and Silicon Telemetry can be used to fully optimize the 3 KPIs at every forward manufacturing/test steps (silicon to server board to datacenter deployment) and reverse steps (silicon health feedback from datacenter for upstream customization/optimization). Telemetry is a powerful capability to generate data at every stack to build supervised learning for AI/ML models. AI and ML capabilities don’t have to be perfect from the start but should improve over time.
In such a complex development in heterogenous integration era to meet the AI demand, siloed engineering development effort no longer works and co-engineering with industry partners is needed. We will also discuss how we worked with industry partners to achieve the goal.