It’s an optical delay-line memory. Early computer memories were acoustic in some manner.
I can’t imagine that the latency of ‘delay line RAM’ would be acceptable to anyone today. Maybe there’s some clever multiplexing that could improve that but it would surely add more complexity that just making more RAM ICs.
Neural net computation has predictable access patterns, so instead of using the thing as a random access memory with latency incurred by waiting for the bit you want to get around to you, I expect that you can load the memory appropriately such that you always have the appropriate bit showing up at the time you need it. I’d guess that it probably needs something like the ability to buffer a small amount of data to get and keep multiple fiber coils in synch due to thermal expansion.
The Hacker’s Jargon File has an anecdote about doing something akin to that with drum memory, “The Story of Mel”.
It’s an optical delay-line memory. Early computer memories were acoustic in some manner.
I can’t imagine that the latency of ‘delay line RAM’ would be acceptable to anyone today. Maybe there’s some clever multiplexing that could improve that but it would surely add more complexity that just making more RAM ICs.
Neural net computation has predictable access patterns, so instead of using the thing as a random access memory with latency incurred by waiting for the bit you want to get around to you, I expect that you can load the memory appropriately such that you always have the appropriate bit showing up at the time you need it. I’d guess that it probably needs something like the ability to buffer a small amount of data to get and keep multiple fiber coils in synch due to thermal expansion.
The Hacker’s Jargon File has an anecdote about doing something akin to that with drum memory, “The Story of Mel”.
http://www.catb.org/~esr/jargon/html/story-of-mel.html