'This hyper-efficient, yet powerful neural processing system, architected for embedded
Edge AI applications, now adds efficient 8-bit processing to go with advanced
capabilities such as time domain convolutions and vision transformer acceleration, for an
unprecedented level of performance in sub-watt devices, taking them from perception
towards cognition.'
'Mr Hehir added, “The development of the second generation of Akida was strongly
influenced by our customers’ feedback and driven by our extensive market engagement.'
Hi Dio,
Am I right in thinking that the AKD1500 doesn't have enough 'grunt' to run 8 bit efficiently. Is this possibly the AKD2000?
And what customers need 8 bit?
Hi FJ,
As we don't have the full details or patent specification, the following is just speculation.
We haven't seen anything to suggest AKD1500 can handle 8 bits. On the other hand, do we have the resources to run two developments in parallel? The release date of Q3 2023 does tie in with AKD1500.
The 8 bits will increase accuracy at the expense of time and power, so the applications will be in fields where accuracy is paramount.
The time penalty will not be such as to take it outside the concept of "real time", but will probably reduce the FPS equivalent rating. I'm not sure how this would affect our compatibility with Prophesee, but I assume we will still have the 1, 2, and 4 bit options.
I suppose the 8 bits will also apply to the model libraries.
This emphasizes the importance of N-of-M coding to provide performance advantages over the competition.
Edit: As
@Boab points out, "scene segmentation" is new. This may dovetail with Luminar's "foveated LiDaR".
www.qualcomm.com
