Hi Tls,@Diogenese I literally have no idea what this means but is is good for BrainChip?
View attachment 19749
François Piednoël de Normandie on LinkedIn: #idc | 10 comments
Intel is now in position it was just before the launch of Conroe. Because competitive pressure, it was forced to push the power envelope and increase… | 10 comments on LinkedInwww.linkedin.com
Transistors have internal resistance, so, when a current flows, heat is generated. If we assume that each operation of a digital transistor switch (CMOS) in any specified geometric chip size uses the same amount of power no matter what chip it is in, then more TOPS means more power. Increasing the clock speed increases power consumption. In an IC, heat dissipates more slowly than it is generated, so there is an accumulation of heat. So to reduce power while obtaining an equivalent outcome means that the number of transistor switch operations must be reduced.
Pumping cooling fluid around and blowing air from a fan through a radiator much like a car's cooling system uses more power.
There are a couple of ways of reducing the number of switch operations, eg, not implementing multiply by zero operations in MAC, or lowering the voltage supply (which has the effect of reducing the signal-to-noise ratio (increasing error-proneness)).
However, these savings are small beer compared to the use of spikes instead of 8-bit or upward bytes to implement a function.
The article is talking about an Intel CPU, not Loihi, so it would be using 64-bit MACs or higher - very high precision, but very power-hungry.
Before Akida, most hardware SNNs used analog ReRAM/MemRistor because it is more closely analogous to real neurons. However, human technology lacks the millions of years of evolutionary refinement which lead to the human brain, and the manufacture of analog neurons lacks the precision necessary to derive consistently accurate outcomes. The performance of analog neurons is too variable, and considering that millions of neurons are included in an SNN, the accumulated errors can be significant. While some attempts to overcome this problem have been made, they generally end up using more power.
As we've discussed before, the discovery by Simon Thorpe's group that the receptor cells in the eye respond more quickly to stronger input signals, so most of the visual information is included in the earlier receptor cell signals. This can also be applied to the pixels of a camera's light sensor.
This lead to N-of-M coding in which only, say, 10% of incoming pixel data needs to be processed by choosing the first 10% of incoming spikes.
A lot of work of a CPU or GPU is image-related in that, apart from any image/video processing, the processor must control the screen display. I wonder if there is a role for Akida there?
I have been pondering Qualcomm and their analog SNN patents for a while. I remember when BRN used to mark them as a competitor by listing them along with Intel and IBM as tech comparisons - you know the checklist ticks against the different categories such as continuous learning and power consumption etc. Then they stopped including them in that list around 2018 or 19. I assumed when that happen that perhaps they’ve bought the carrot and would be a future customer. There’s been a lot said about the inherent instability in analog SNNs. Do we think that somehow Qualcomm have solved this? Not likely, unless they’ve done so somewhere in the last 18 months and patents will soon be published to reflect that. That would be a pretty big deal considering academia world wide has yet to bare such an accomplishment. Not to say it’s impossible for Qualcomm but you would think they would need to have harnessed the crème of Nueromorphic talent from academia to do so. So perhaps it’s then possible that Qualcomm’s Snapdragon platform includes a bit of our magic to make it all work, I think so anyway, just rambling on a bit haven’t done so for a while - thanks to all the great contributors to this forum I haven’t felt the need to. Cheers to all.
Back in Akida Research HQ (Perth) after a great break to see family/friends in Canada. 
. While enjoying the broadcast of the US Open and various NFL games I would use the ad breaks to catch up on the dot joining happening on TSE. I can't explain to you the pleasure of hearing the words "neuromorphic computing 5x more intelligent then current voice AI" come blaring through the TV screen and the smile it brought to my face
. Mercedes are pushing their advertising hard. I got to hear those words and see that ad on multiple occasions
. 
