Magnetic tunnel junction-based computational random-access memory | Hacker News Reader

Magnetic tunnel junction-based computational random-access memory

arxiv.org

34 pointsPaulHoule6 comments

I first heard of computational RAM some time around 1999-2001 [edit: in the context of parallel rendering]. Looking through the bibliography here most of their computational citations are from 2011-2017, which confused me at first.

I think the deal with this paper is not what, but how.

    In this work, a CRAM array based on magnetic tunnel junctions (MTJs) is experimentally demonstrated.
    First, basic memory operations as well as 2-, 3-, and 5-input logic operations are studied. Then, a
    1-bit full adder with two different designs is demonstrated. Based on the experimental results, a
    suite of modeling has been developed to characterize the accuracy of CRAM computation. Further 
    analysis of scalar addition, multiplication, and matrix multiplication shows promising results.

So it's a funky kind of memory cell. Which they then...

    These results are then applied to a complete application: a neural network based handwritten digit classifier,

So we have again arrived at the point in the R & D epicycle where every research paper is linked to the buzzword of the day to attract the very short attention spans of investors.

No, CRAM has been of interest for neural networks for a long, long time. I find the “buzzword” analysis to be infuriating because people have been developing neural systems for decades: as annoying as those NIST digits (a facile response to so many papers is: “has anyone ever asked if it would work for some other data set?”) but Yan LeCunn delivered a handwritten digit recognizer to the US post office in the early 90s, I worked on a neural search engine for patents 10 years ago, etc.

If you are annoyed by the people who pivoted from NFTs to LLMs, I am annoyed even more.

CRAM is not limited to AI. I first saw it demonstrated as a distributed 3D graphics engine (though at the time it was more like vector graphics) where each chip was responsible for a different section of the display. Closer to ray tracing than neural networks.

... and lots of other things too. One early application of content-addressable memory was STARAN

https://en.wikipedia.org/wiki/STARAN

which was used in air traffic control systems to project the trajectories of all the aircraft being tracked and test if any are at risk of collision.

What is Computational RAM (CRAM)?

> CRAM performs logic operations directly using the memory cells themselves, without having the data ever leave the memory.

Sounds like an interesting new approach to computation if we can make this work at scale.

Does that mean no registers? Is there still a need for L1/2/3 caches?

Reply on news.ycombinator.com