Learning Opportunity: Hotchips free for all

The Hotchips conference is one of the key events to understand how computing hardware is evolving. As hardware advances, software races to exploit it. The presentations and videos from the Aug 21 Hotchips33 conference are online and worth a look. If you are involving with AI, Machine learning, some good tutorial sessions from Google, Intel, Facebook etc Some datapoints that I thought were interesting:

• Top of my list was the DeepSpeed DL training and inference optimization library from Microsoft which can accelerate inference models up to 6x. Best part is it is available on Github!
• The Google session on the MLPerf training benchmarks show ML results are outstripping Moore's Law with a 6.8x -11x improvement
• If you have any doubt about the importance of ethical AI in our current environment, get sense of the scale of AI and ML training at Facebook
• Google new TPU-V4 provides some impressive results - Pods with 4096 chips for 1.1 exaflops of mixed precision?
• Esperanto use of RISC processors (is it back?) for ML recomendations: T-SoC-1 has over a thousand RISC-V processors on a single TSMC 7nm chip.
• The Cerebas wafer scale chip targeting at accelerating AI: 46,225 mm2 silicon, 2.6 Trillion transistors 850,000 AI optimized cores 40 Gigabytes on-chip memory 20 Petabyte/s memory bandwidth. Changing the game to give Nvidia a run for the money.

You have the hardware specific sessions from Intel, AMd, Nvidia .... But it was great to see some of the emerging technologies suppliers I am tracking like Mojo Vision (AR Contact Lenses) and Skydio (autonomous flight drone) presentations