Why Akida 2?
Sean Hehir mentioned that Akida 1 has narrow application, so what does Akida 2 add?
- 8-bit weights and activations;
- Vision transformer;
- Long skip;
- TENNs.
What do these features facilitate? (
https://brainchip.com/akida-generations/ )
- Audio Processing, Filtering, Denoising for Hearing Aids, and Hearables
- Speech Processing for Consumer, Automotive and Industrial Applications
- Vital Signs Prediction & Monitoring in Healthcare
- Time Series Forecasting for Industrial Predictive Maintenance
- Video Object Detection and Tracking in Consumer, Automotive and Industrial
- Vision, LIDAR Analysis for ADAS
- Advanced Sequence Prediction for Robotic Navigation and Autonomy
These capabilities are critically needed in Industrial, Automotive, Digital Health, Smart Home, and Smart City Applications.
We’ve heard a little about denoising earphones/hearing aids.
We already knew about key word spotting , but speech processing involves the temporal element and long skip (to preserve the context of the speech).
Healthcare involves reading physiological signals, X-ray and ultrasound interpretation, etc.
Predictive maintenance from vibration analysis.
Video object detection was within Akida 1’s capability, so the tracking is new – this would have needed to be done by an associated CPU for Akida 1. This is for consumer, automotive, and industrial applications. It would be especially advantageous for EVs as it reduces the CPU power drain.
Vision, lidar analysis for ADAS – Akida 1 does vision/lidar classification, so “analysis” is an added feature, again relieving the CPU of the task. In particular, this may relate to object identification and tracking, possibly utilizing vision transformer and long skip to keep track of an object’s movements. Similar considerations apply to robotic autonomy.
So, all-in-all, I think that the Akida 2/TENNs improvements bring significant power and latency benefits for EVs.
Unfortunately, as far as we know, there is no Akida 2 silicon. Still, using Akida and or TENNs models with the corresponding simulation software would bring substantial advantages to battery-powered devices including EVs, because Akida models are much more compact than competing tech models, and TENNs models are more compact again.
We know Valeo, Mercedes and others are developing SDVs.
We know we've been working with Valeo and Mercedes for a few years. Even if there has been insufficient testing of Akida 2 for safety-critical applications, I think there would be scope for Akida 2's long skip to be used in speech recognition in the infotainment system in the not too distant future.
TSMC founder criticizes Intel's foundry pivot; says AI focus was crucial (Image source: Intel)
and all the laughter post throughout the years 
