Uniquely scalable neural network Ai chip using FFT techniques
Using the Teensy 4.1 for high level control and development ease, this
Ai approach breaks the computational bottlenecks
associated with real time convolution, with the additional benefit
of radically reduced software overhead. Most
demanding Ai tasks such as frequency domain matrix multiplies are
reduced to less than a dozen
machine instructions.
We concentrates on performing the
algorithms using the efficiencies of the FFT.
Applications using time domain techniques such as the
A100 and newer H100 from Nvidia increase in operations by a N squared
function. These same applications using FFT
techniques increase in operations by N x Log (N), a
exponential reduction.
Multiple chips can be cascaded for higher radices, two
cascaded chips can perform a 1024 radix, this gives a 1
million point transform in just two passes.
This chip is being designed using macro cells and super cells. The end customer can use
the approach as a silicon core and surround the chip with
proprietary circuitry to further enhance the application
and reduce his overall application hardware
Interesting. Did you design the system yourself?
This DSP-24 chip apparently was designed around the year 2000. The company DSP Architectures doesn't seem to be alive anymore.
Just curious why you used this particular DSP chip.
Interesting. Did you design the system yourself?
This DSP-24 chip apparently was designed around the year 2000. The company DSP Architectures doesn't seem to be alive anymore.
Just curious why you used this particular DSP chip.
Paul
Hi Paul. Yes, you are right on all counts. When those twin towers came down, the DoD pulled all of our funding. I owned DSP Architectures Inc, myself and a handful of great Engineers designed the DSP24.
I believe that the exploding Ai market is a "killer app" for this chip architecture. I will be posting my progress on this Teensy forum. A couple links below, if you want to explore further.
I believe that the exploding Ai market is a "killer app" for this chip architecture
Hi Michael, reading your statement: do you see a potential revival of this particular chip architecture? Or perhaps even an update of the DSP24 chip?
Do you have a specific area of the AI market in mind?
The absolute King of Ai is Nvidia. No large cap company has ever done what they just did in the stock market, see image below:
The only company with a chance of overtaking Nvidia is probably Tesla. It would be silly for anyone to get near these monsters, they have entered the fight of all time.
I'm positioning to be out front in the niche called edge computing. The DSP24 was a proof of concept chip that I'm using to show the path forward.
Reply With Quote
Originally Posted by joepasquariello
