Challenges Emerge for In-Memory Computing

For each layer of the network, both the existing weights and the elements of each training example are loaded into the processor’s registers, multiplied, and the result written back out to memory. Low precision, analog memory, high accuracy

In an IEDM presentation, Jeff Welser, vice president and lab director at IBM Research Almaden, noted that high computational precision generally is not needed for neural network calculations. However, data storage applications require digital values — a device is either on or off — and typically use strong enough SET and RESET pulses to create or remove a strong conductive filament.

Source: semiengineering.com