SRAM
A fast type of computer memory that sits directly on the processor die, holding data without needing constant refreshing.
Computers use two main kinds of working memory. DRAM is the cheap, dense sort that goes in the sticks you plug into a motherboard, and it has to be electrically refreshed thousands of times a second to hold on to anything. SRAM is the expensive, fast sort built directly into the processor itself, and it holds its contents as long as the power is on. Every processor already contains some SRAM as cache.
The reason SRAM keeps coming up in AI conversations is distance. Fetching a model’s numbers from separate memory chips costs time and a great deal of energy, often more than the calculation itself. Several chip startups, Fractile and Cerebras among them, are betting on keeping far more of the model in on-die SRAM so it never has to travel. The catch is that SRAM takes up a lot of chip area for each bit stored, so you get speed at the cost of capacity and price.