Skip to content

History & Evolution of Inexact Computing

How did the computing world transition from the rigid demand for 100% mathematical perfection to intentionally designing circuits that make calculated mistakes?

⏳ Timeline: The Century-Long Journey

timeline
    title The Evolution of Arithmetic Exactness in Silicon
    1945 - 1980 : The Exactness Era : Von Neumann & Turing demand 100% deterministic logic. Babbage engines and early mainframes treat a single bit flip as catastrophic hardware failure.
    1985 - 2005 : The Golden Age of Moore's Law : Dennard Scaling allows chips to get faster and cooler automatically every 18 months. Exactness is "free" because power consumption is low.
    2005 - 2010 : The Power Wall & Dark Silicon : Dennard scaling collapses! Clock speeds stall at ~4 GHz. Chips generate too much heat to power all transistors simultaneously ("Dark Silicon").
    2010 - 2018 : The Birth of Inexact Computing : Prof. Krishna Palem and pioneers prove probabilistic and approximate arithmetic cuts 70%+ power. First approximate adders (ACA, LOA) and multipliers (DRUM, RoBA) appear.
    2018 - Present : The AI & Edge Computing Revolution : Large Language Models (ChatGPT) and edge sensors explode. Chips cannot afford exactness tax; approximate TPUs and stochastic accelerators become mainstream.

🏛️ Phase 1: The Quest for Absolute Perfection (1940s – 1990s)

In the dawn of computer science, early computers were built for ballistics calculation, nuclear physics simulations, and accounting. In those domains, precision was non-negotiable: - If a bank balance is off by \(\$0.01\), people lose trust. - If an orbital trajectory calculation drops a carry bit, a spacecraft misses its target.

Because early vacuum tubes and relays were prone to random hardware faults, computer pioneers (John von Neumann, Claude Shannon, Richard Hamming) spent decades creating Error-Correcting Codes (ECC), triple-modular redundancy, and strict synchronous timing to ensure that 1 + 1 always equaled 2, zero exceptions.

💥 Phase 2: The Collapse of Dennard Scaling (2005)

For thirty years, Dennard Scaling was the magic engine of the tech industry: as transistors shrank, their power density stayed constant. You could pack \(2\times\) more transistors, run them faster, and the chip wouldn't melt.

Around 2005, Dennard scaling hit the laws of quantum physics: 1. Voltage Scaling Stalled: Transistors became so small that electrical current began leaking through the microscopic gate oxides even when turned off. 2. The Thermal Wall: If chip designers kept increasing clock frequencies, processor dies would reach temperatures hotter than a rocket nozzle! 3. The "Dark Silicon" Dilemma: In modern sub-7nm chips, over 50% of the silicon must remain powered off ("dark") at any given moment to prevent the processor from burning itself out.

+-------------------------------------------------------------------+
|                     THE POST-2005 SILICON CRISIS                  |
|                                                                   |
|   Transistors keep shrinking ──> But cannot be powered all at once|
|   Result: We cannot make exact circuits faster by brute force!     |
+-------------------------------------------------------------------+

💡 Phase 3: The Paradigm Shift (Palem & Inexact Computing)

In the late 2000s, researchers like Prof. Krishna Palem (Rice University / NTU) and leading VLSI architects asked a revolutionary question:

"Why are we spending 70% of our precious battery power to guarantee 100% precision for applications that don't need it?"

They observed that the fastest-growing computing workloads—digital photos, video streaming, music synthesis, and machine learning—had something in common: - The human eye cannot tell if a background pixel is 128 or 129. - The human ear cannot distinguish microscopic phase jitter in a sound wave. - A neural network doesn't care if an image weight is \(0.4501\) instead of \(0.4500\).

By intentionally relaxing the requirement for 100% precision, engineers discovered they could prune carry chains, eliminate Wallace-tree compressors, and drop operating voltages, achieving up to \(3\times\) to \(5\times\) energy efficiency improvements.

📚 Primary Literature & Historical Citations

For students and historians studying the founding papers of inexact and approximate computing: