Hi JB,
Quadric make what they call a "Supercomputer" which integrates the functions of CPU/GPT/AI accelerator using a modified CNN for image classification.
Its basic unit seems to be their Vortex ALU.
It won't be used for doorbells.
The White Paper 2019
May, 2019
https://www.quadric.io/post/the-white-paper
we founded Quadric to build a product that brings server-class performance to the edge.
View attachment 2723
By 2015, researchers were proving classification results on the Imagenet challenge that exceeded human error rates. One such advanced CNN network architecture, RESNET50 strikes a balance between total computational network complexity and error rate.
View attachment 2724
####################################################################################
https://brainchipinc.com/wp-content...brief_6-How-BrainChip-is-Changing-AI_v1.2.pdf
The Akida Event-Based Neural Processor
The Akida event-based neural processor is a fundamentally different approach that breaks the linear relationship between high power consumption and performance seen in traditional accelerators. The Akida processor is 10x to 30x more energy-efficient than its nearest competitor for inferencing on industry-standard benchmarks such as MobileNet and Google Keyword Spotting DNNs, and is easy to use. Trained on MobileNet’s Imagenet 1000 data set, the Akida neural processor can classify all 1.2 million images, and 1,000 classes, at 30 frames per second within a power budget of just 156 milliwatts in 28nm, compared to several watts for a Google Edge TPU. Audio keyword recognition using the Google keyword database runs at an extremely low power of 150 microwatts.
30 frames per second in 156 mW = 192 frames per second per Watt. (This is the 2019 Akida, not the improved 2021 commercial version - to be fair, the Quadric figures are from a 2019 white paper). So Quadric is in the same ballpark, but Akida has made home base, while Quadric just making 3rd.
The main difference is that Quadric is a Frankenstein amalgam of CPU, GPU and AI accelerator. Their AI accelerator uses ResNet50, an improved CNN arrangement which enables skip connections when intermediate layers are not required.
View attachment 2731
Also remember Akida's performance at key word spotting:
View attachment 2726
IPS = inferences (identified words) per second