Sandia's analog memory claims 100 times the precision of an unnamed state of the art

Somebody at a camera-module company reads this and goes looking for a datasheet. There isn't one. What is on the record is two multipliers, a mechanism and a materials stack.

Start with the multiplier that is hardest to use. Fuller said ETCRAM "is able to achieve 100 times higher precision than existing state-of-the-art technology," and separately "at least three orders of magnitude greater dynamic range" ^[3]. Three orders of magnitude is a factor of 1,000 ^[9]. The precision figure needs a conversion before an engineer can do anything with it. If 100 times precision means 100 times as many distinguishable levels in a cell, that is about 6.6 extra bits, because log2(100) is 6.64 ^[10]. Six and a half bits on top of what, the account does not say. No competing device, process or part number is named ^[11].

Then the mechanism. Electrochemistry is slow, and the fix described is self-heating: micro-bursts of heat triggered as the electrical pulse arrives, which the lab says speeds the device up and locks in precise analog values without pulling large power ^[6]. Alec Talin's framing is a battery. "You can charge a battery halfway and then stop charging and use that energy. Think of the state of the battery as a memory. It stores that state," Talin said ^[5]. How long the state holds is the first thing a designer asks about an analog cell. Retention is not in the account, and neither is endurance, write energy, array size or a date ^[11].

The energy argument does not line up with the application. The Interesting Engineering writeup cites federal estimates that US server electricity consumption could reach roughly 800 billion kilowatt-hours a year by 2050 ^[7], and describes the payoff as a security camera analyzing video on its own sensor ^[8]. A camera sensor is not a server, so a design win in edge silicon would not move that total. The account does not tie the projection to a named agency or report ^[11].

Which part is the 100x measured against, by name. How many bits stay resolvable after a realistic count of write cycles and after storage at temperature. Has anyone written and read a full array. None of the three is answered in the public account ^[11]. The idea that precision and dynamic range are the specific gap keeping analog in-memory compute out of shipping edge silicon is Fuller's description of his own device's advantage; the account names no rival device and gives no yield or integration detail ^[11].