Building AI Chips that Deliver | Kisaco Research

As AI & ML chips proliferate, the focus is moving to deployment and practical applications. This change brings a corresponding shift in the technical problems chip makers and their customers need to solve. Do the devices perform as expected under real-world conditions? What support do software development teams – both at the IC maker and systems integrator – require? Can devices adapt as software loads change and the devices themselves are subject to aging effects?

The pioneers of the current generation of AI and machine learning devices faced a fundamental problem: how to build architectures incorporating many, many processing nodes, that deal with often sparse data sets and deliver results while achieving any level of resource efficiency.

The world has moved on, and today’s AI/ML chips now deal efficiently with their target workloads. But echoes of the early efficiency and performance problems remain. Chip and system performance metrics are often dominated by infrequently occurring long-tail phenomena which may not show up at first sight as bugs. Interactions between on-chip hardware blocks are not predictable – at least in the sense that traditional designers understand. Hardware and software need to be designed and optimized in parallel. The sheer size and complexity of many of these new devices makes it impossible to simulate or emulate every aspect of their performance before tape-out – and in the case of real-time and cyberphysical systems, to predict every possible interaction with the real world.

This panel-style discussion will address these issues from a number of standpoints:

  • What are the main deployment challenges at the technical level?
  • What is required from this new generation of chips to allow them to be integrated successfully into target systems?
  • Looking beyond systems integration, what are the requirements when the chips are deployed in the field?
  • How can IC design teams ensure their chips attain maximum performance?
  • What tools and methodologies are available to support them?
Speakers: 
Gajinder Panesar
Rangan Sukumar
Dennis Abts
Ty Garibay
Date / Time: 
Wednesday, September 1, 2021 - 10:00am
Call to Action Label (Upcoming): 
Register Today
On-Demand Option: 
Simple Vimeo video (no registration required)
On-Demand Registration is the same as the registration URL: 
0
Call to Action Label (On-Demand): 
View Webinar
Event Type: 
Webinar Type: 
Sponsor List: