John Carmack sinniert über die Verwendung einer langen Glasfaserleitung als L2-Cache für das Streaming von KI-Daten – Programmierer stellt sich Glasfaser als Alternative zu DRAM vor
https://www.tomshardware.com/pc-components/ram/john-carmack-muses-using-a-long-fiber-line-as-as-an-l2-cache-for-streaming-ai-data-programmer-imagines-fiber-as-alternative-to-dram
15 Kommentare
data isn’t stored in the fiber at all. It’s a continuous medium not discrete. data moves through the fiber. There are no individual addresses to probe that aren’t in constant flux
it would have to be a very long fiber… let’s see 1 tb per second is data through fiber approx, let’s say the light moves at 0.5 c through the medium, 150,000,000 m/s, or in 1 ns light moves about a foot in computer chips, i think that’s one bit, per foot. please correct me. thanks!
Sounds like the first memory devices IBM invented, a very long coiled wired and they would twitch the input, the twitch would propagate through the wire until it got to the end of the coil and then the output was fed back into the input.
Return of the delay line!
[https://en.wikipedia.org/wiki/Delay-line_memory](https://en.wikipedia.org/wiki/Delay-line_memory)
That’s a delay line, not addressable memory. They are different.
Let’s all just take a moment to consider that maybe Carmack was high as a kite. Cache is useful if it’s addressable, and continually moving light is not, so far as I’m aware.
Word salad title was generated by the fiber-contained data as well?
John Carmack like the guy who made doom? Hes gotten into ai garbage now? Get fucked creep
I can’t wait for Civvie to give this guy an even longer, more abstract nickname for this.
RA in DRAM stands for Random Addressable. Fiber is more akin to FIFO buffer.
Fun Fact : On his honeymoon his wife demanded he not take a computer or device with him. During a walk on the beach he came up with what ended up being ID’s MegaTexture technology that they used for years. He went back to his hotel room and wrote out the code by hand on paper.
Somebody call Civvie
Memory access patterns for transformer models are very regular and periodic but high bandwidth. The memory access patterns to load the full weights of a model into memory for each token are exactly the same for each token (mostly) so I could see how, if you measured how quickly the processor theoretically churn through the model parameters, you could loop those parameters through the optics to get to the cpu at exactly the right time during each token cycle.
Dear John Carmack: Please don’t invent the fiber optic rationing system so that Grok reserves 90% of consumer bandwidth. You could take up knitting or something.
I will always say carmack is a genius like probably our gens Einstein. You should look at all his out of the box thinking. It’s phenomenal.