Marvell just flexed some seriously efficient SRAM tech at their analyst day, and the numbers make standard solutions look pretty rough. Their 2nm IP burns 80% less power than typical 256K implementations, takes up 37% less space, and cycles 22% quicker. The layout is more rectangular, which apparently makes it way easier to cram into packed SoCs without everything turning into a nightmare.
When stacked against competitors, the custom design uses half the area at matching bandwidth levels, cuts standby drain by two-thirds, and pumps out 17 times more bandwidth per square millimeter. This matters because memory scaling has been lagging behind logic improvements for years. TSMC's N3 SRAM cells are basically identical to N5 despite logic...