2022 BRN Discussion, page-36

Hi Barrel,

I get why you would think this from reading the announcement but I still disagree. When they talk about "improved performance and lower power consumption", it makes no sense to me that the improved performance is referring to accuracy. I think the performance metric must be talking about speed. Maybe the chip now computes faster or can handle higher bandwidth or shares tasks better across cores etc but it shouldn't impact the accuracy.

The hardware should have no impact on the model accuracy so long as it is still only working with a maximum of 4 bits. The Akida model would be trained on say Tensorflow and converted to the Akida model. It is then uploaded to the chip and ran. The Akida model would lose accuracy here because the resolution drops to 4 bits. However, there is no other way in which the hardware should influence the model unless it allows for other software tuning / customisation of the Akida models. But even if this was the case this would just mean the Akida model was tuned better, which is software related. Machine learning accuracy is tied to the neural network model and the learning algorithm, not the hardware it runs on.

Judging by the description in the announcement of "better layout and minor design changes" I personally doubt the accuracy was affected. I think the design was simply optimised further for speed and improved power performance.

Pure speculation, DYOR