>>106570952
It's not "the future" but it probably has a place to run AI on low-power devices (drones, cameras)
By AI I mean actually useful stuff like object detection, not LLMs

>>106570967
Isn't biggest problem facing lot of AI energy consumption with biggest source being fetching weights from memory? Constantly fetching data millions of times in a second is energy intensive so if they're able to bypass that by doing calculations within memory you could reduce energy needs by 50%

>>106570952
No, it's a retarded idea. As you go up in frequency, signal to noise ratio degrades. The reason you cannot go higher in frequency is because if you go any higher the gates will sometimes misidentify a 0 as a 1 or a 1 as a 0.
If you can barely distinguish a 0 from a 1 how are you going to distinguish a 0.000001 from a 0.000002? It's an idea that seems interesting when you know a little but once you learn a little more it becomes pretty much useless.
With digital you can run the whole system at the edge of chaos because of signal reconstruction. If you can tell a 0 from a 1 then you can make the 0 a stronger 0 or the 1 a stronger 1. With analog there are no pre-set values to snap the signal into, you can only add noise to the signal, never remove it.

>>106571016
Yes, as I understand it, that is more or less what Groq does. Each card has like 256MB of memory and you need dozens or hundreds of cards to serve a model, but if you have a thousand concurrent users it ends up being more economical than using GPUs.

>>106571290
>As you go up in frequency, signal to noise ratio degrades.

Analog Compute in memory actually aims for low precision and to create models that can work with less quantization.


IBM is actually attempting to use miniaturized ADC that are parallelized and instead convert time domain instead of actual voltage.


Groq chip doesn't do memory in compute but instead aims for Less movement.

You can multiply matrices with photons, that's much cooler, literally and figuratively.