Why We Never Replaced the Von Neumann Architecture (And Probably Never Will)
Per Stenström explains why the von Neumann architecture still dominates—and whether a worthy successor exists.
Remember learning that a computer consists of a CPU, memory, and I/O devices? Congratulations, you know the von Neumann architecture. Professor Per Stenström from Uppsala University asks: why, after 70 years, haven't we come up with anything fundamentally new?
Spoiler: attempts were made. Harvard architecture, dataflow machines, neuromorphic chips—all either stayed in labs or quietly retreated to specialized niches. As Stenström puts it, "we tried, but von Neumann turned out to be that old friend you're comfortable with, even if he snores."
The main issue—legacy. All our software is built for the classic model: sequential execution, separate code and data. Switching to something new would require rewriting everything—from operating systems to microservices. And who in their right mind wants to rewrite '60s legacy?
Still, Stenström notes the bottleneck remains the "von Neumann bottleneck"—the data transfer speed between CPU and memory. That's where the battle happens: caches, branch prediction, multithreading. But the foundation stays the same.
METABYTE studio comment: The von Neumann architecture is like jQuery: everyone knows it's time to move on, but the codebase is already so huge that it's easier to drag it to retirement. At METABYTE we're not against legacy, but if your project starts to lag—maybe it's time to look into FPGAs or at least decent caching?
NEXT STEP
Liked the approach?
We apply the same principles to client projects: AI, automation, products that don't die after launch.