The lack of investment in more production capacity for RAM is based on a roughly 3-year horizon for this insane extra AI demand.
Creating workable consumer-grade alternatives with delay line memory of all things would take longer than that, and the market would collapse the moment AI demand for RAM dried up. This is one of those things that is theoretically possible but due to both technology and market conditions will absolutely not be a thing.
I think that this is intended not to replace DIMMs in PCs, but to replace HBM for AI use. If you’re doing neural net computation, you have very predictable access patterns, so you can store your edge weights such that the desired data is showing up at just the right time.
https://en.wikipedia.org/wiki/Delay-line_memory
The lack of investment in more production capacity for RAM is based on a roughly 3-year horizon for this insane extra AI demand.
Creating workable consumer-grade alternatives with delay line memory of all things would take longer than that, and the market would collapse the moment AI demand for RAM dried up. This is one of those things that is theoretically possible but due to both technology and market conditions will absolutely not be a thing.
I think that this is intended not to replace DIMMs in PCs, but to replace HBM for AI use. If you’re doing neural net computation, you have very predictable access patterns, so you can store your edge weights such that the desired data is showing up at just the right time.
It’s not that we don’t know what it is, it, again, is just INCREDIBLY FUCKING STUPID.