Long before the term "artificial intelligence" became a household phrase, engineers and mathematicians wrestled with a fundamental bottleneck in computing: the physical separation between where data is stored and where it is processed. That divide, known informally as the "von Neumann bottleneck" after the architect whose 1945 design defined modern computers, has haunted the field for eight decades. Now, a wave of research into memristor-based analogue computing is mounting what may be the most credible challenge yet to that entrenched architecture.
A study published in Nature outlines strategies for achieving high-accuracy analogue computation directly within memory arrays, using memristors — resistive devices whose conductance can be tuned to represent weighted values — to perform the matrix multiplications that underpin nearly every modern AI workload. The approach belongs to a broader class called computing-in-memory, or CIM, which traces conceptual roots back to proposals from the 1960s and gained renewed urgency as deep learning's appetite for energy and bandwidth ballooned in the 2010s.
What makes the current moment historically significant is the convergence of materials science, fabrication precision, and algorithmic compensation techniques that together begin to overcome analogue computing's traditional nemesis: device variation and noise. Early analogue neural network hardware, including pioneering work at Bell Labs and later DARPA-funded projects in the 1980s and 1990s, foundered precisely on these reliability problems. Digital silicon ultimately won that era's competition because it was predictable, programmable, and scalable.
Today's memristor research inherits that lineage but arrives equipped with decades of learned lessons. Researchers are now pairing imperfect analogue devices with sophisticated error-correction and calibration schemes, effectively letting software compensate for hardware imperfections — a hybrid philosophy that echoes how early transistor radios managed noisy components through circuit design.
Whether memristor arrays can ultimately displace conventional accelerators for AI inference remains an open question. But the trajectory is clear: the field is moving steadily toward hardware that thinks and remembers in the same place, closing a loop that von Neumann's original blueprint left deliberately open. For historians of computing, that would represent not merely a technical upgrade, but a genuine architectural inflection point.