From IEDM'22:

TSMC's SRAM cell size is only 5% smaller at 3 nm.

#industry #sota #ai

@yigit And do you know how much smaller would be a memristor?

(Or maybe how much smaller is a crossbar array of comparable storage capacity? Or maybe another comparison is relevant?)

@BellecGuill

I can check the SOTA but <100nm^2 is a reasonable estimate for cell area. Smallest memristor cell I worked has a footprint of 8x20nm^2, it was a single-bit device.
For inference applications where the device can be programmed with computer in the loop via iterative programming, 11-bit is the highest capacity I know (SRAM stores 1-bit)

For training where the device needs to be programmed by application of single pulse without verification, 3-4 bits resolution is already available.

@yigit @BellecGuill
Excellent, so with a conservative estimate 100 nm2 and 4 bit per memristor, it gives a factor x3200 between asymptomatical SRAM area and the memristor equivalent.

I had no idea the scaling was so massive !