← All publications

Working Paper

FLUID: A New Way to Think About Computing

An Introduction to Token-Based Dataflow Architecture

The architecture of modern computers was designed in 1945. For eight decades we've optimized the sequential von Neumann bottleneck instead of replacing it. FLUID asks: what happens when we rethink computation from scratch?

This paper introduces FLUID — a dataflow architecture where self-describing tokens flow between specialized Processing Elements. Parallelism emerges naturally, security lives in the hardware, and computation becomes communication.

Author Hannes Lehmann / sistemica GmbH
Date February 2026
Download English (34 Pages)

Tokens, Not Instructions

256-bit self-describing packets replace the fetch-decode-execute cycle

Modular Architecture

Extensible with analog, optical, or quantum Processing Elements

Computation = Communication

Data flow is the program — no registers, no program counter

Deterministic Execution

Same inputs, same graph → same outputs in the same number of cycles

Why do we need a new kind of computer?

Imagine a factory where every worker must ask permission from a single foreman before doing anything. Worker A finishes their task and sets down their output. Before Worker B can pick it up and continue, they have to wait for the foreman to notice that A is done, walk over, check the paperwork, and give explicit instructions to B. Meanwhile, Workers C through Z stand idle at their stations, even though they have everything they need to work and their tasks don't depend on what A and B are doing.

This is how your computer works today. The architecture of modern computers was designed in 1945 by John von Neumann, at a time when memory was extraordinarily expensive, transistors didn't exist yet, and the idea of parallel processing was pure science fiction. The solution von Neumann proposed was elegant for its time: fetch one instruction from memory, execute it, fetch the next instruction, execute it, and repeat forever. This sequential model made sense when building a computer meant wiring together thousands of vacuum tubes by hand.

Eighty years later, we still use this basic design. We've made it dramatically faster through billions of transistors, multiple processor cores, and clock speeds measured in gigahertz. We've added caches, branch predictors, speculative execution, and countless other optimizations. But the fundamental bottleneck remains: a single stream of instructions flowing through a central processor, one after another, with elaborate machinery trying to squeeze more work out of each cycle.

This is the von Neumann bottleneck, and its consequences touch every aspect of modern computing. Your phone gets hot when running AI models because the processor is working furiously to push data through a pipeline designed for sequential math, not the massively parallel operations that neural networks require. Security vulnerabilities like Spectre and Meltdown exist because processors try to guess what instruction comes next and execute it speculatively, creating side channels that attackers can exploit.

These problems have been understood for decades. John Backus identified the von Neumann bottleneck by name in his 1977 Turing Award lecture. Since the 1970s, researchers have pursued fundamentally different architectures — dataflow machines at MIT and Manchester, the Inmos transputer, the Connection Machine. Each approach solved part of the problem. None addressed the full challenge: an architecture that makes parallelism natural, security intrinsic, and general-purpose computing practical — all from the same set of design principles.