Researchers are looking beyond digital computing

But shuttling data between the processor and memory eats up time and energy, particularly for data-hungry artificial-intelligence (ai) models. An analysis by Amir Gholami and colleagues at the University of California, Berkeley found that in the past two decades, processor performance has tripled every two years, while memory access speed has increased only by about half (see chart). This means processors blaze through calculations faster than data can be fed from memory, creating a “von Neumann bottleneck”. This has led some to wonder if it is time for a new architecture. Every engineer, and every reader, carries with them the proof that such a thing is possible.