“We demonstrate that the proposed SSA framework achieves close to baseline ANN accuracy for image classification tasks with ViT-Small model, with over 1.8× estimated gains in energy efficiency. Furthermore, an FPGA implementation of an SSA block shows 48× lower latency while consuming 15× less power than GPUs. Overall, this work supports the implementation of Transformer-based models for energy and resource constrained applications, including mobile devices in 6G.”
The above research did not involve a consideration of Brainchip AKIDA technology and did not have the benefit of Brainchip’s patent protected TENNS.
Notwithstanding the researchers are engaged in proving out the utility of using Spiking Neural Networks for Vision Transformers (ViT) and in so doing provide independent collaboration of the claims made by Brainchip.
My opinion only DYOR Fact Finder
BRN Price at posting:
35.0¢ Sentiment: Hold Disclosure: Held
(20min delay)