Fun keynote from @SkydioHQ at #hotchips last night.
Interesting learnings for architects designing compute SoCs for edge devices in #autonomous systems & #Robotics
Last two were the most revealing realizations for me (1/n)
Interesting learnings for architects designing compute SoCs for edge devices in #autonomous systems & #Robotics
Last two were the most revealing realizations for me (1/n)
https://twitter.com/SkydioHQ/status/1429458489735254016
Autonomous algorithms are still evolving.
While acceleration of key workloads is desired, it is general compute horsepower which will provide the needed flexibility to program solutions for next world challenges. (2/n)
While acceleration of key workloads is desired, it is general compute horsepower which will provide the needed flexibility to program solutions for next world challenges. (2/n)
Good software abstraction of foundational building blocks allows engineers to iterate faster with different sophisticated algorithms.
(3/n)
(3/n)
Heterogenous compute requirements get mapped onto different compute units currently. (4/n)
Deep Learning at this edge is NOT limited to INT8 or lower precision operations.
Some layers still need floating point precision. (5/n)
Some layers still need floating point precision. (5/n)
NonLinear Optimizations which are the core of many #Robotics algorithms still get deployed on CPU.
There is an opportunity to offload these fairly compute intensive steps onto accelerators. (n/n)
There is an opportunity to offload these fairly compute intensive steps onto accelerators. (n/n)
• • •
Missing some Tweet in this thread? You can try to
force a refresh
