The Energetics of Computing in Life and Machines, Reviewed. What Computation Costs in Cells, Brains and Chips
The Energetics of Computing in Life and Machines
The SFI Press volume edited by David Wolpert asks what computing must cost in energy, in cells and chips. What its chapters argue and where the limits sit.
The Energetics of Computing in Life and Machines asks how much energy a computation must cost, and whether cells compute more efficiently than chips. David H. Wolpert, Chris Kempes, Peter F. Stadler and Joshua A. Grochow edited it for SFI Press in 2019, as the fourth volume of the Santa Fe Institute’s Seminar series. Its fifteen chapters bring together nonequilibrium physics, computer science theory and biochemistry, with the stated aim of a modern, nonequilibrium thermodynamics of computation.
In short, the book replaces the single famous number of the field, Landauer’s cost of erasing one bit, with a fuller account of where the energy of computing goes. That account sets a floor under any physical system that computes, and it applies to brains as much as to chips.
Where to get it
SFI Press offers the PDF free of charge on the book’s page, and posts the preface and an excerpt of every chapter on its site. This review draws on those public texts, and every claim below is attributed to the chapter that makes it. The book is also sold in paperback, hardcover and ebook on Amazon. As an Amazon Associate I earn from qualifying purchases.
What the book argues
The preface opens with two energy budgets. About 5 percent of all energy consumed in the US goes to running computers, the editors write, citing a Lawrence Berkeley National Laboratory field study, and similar figures are reported for Europe. The human brain, which the preface calls a computer, uses some 10 to 20 percent of the calories a person consumes. The editors draw one conclusion from the second figure. Natural selection acts on the brain’s combined energetic and computational efficiency, so the thermodynamic cost of biological intelligence is a question for evolutionary biology.
| Chapter | Authors | What it contributes |
|---|---|---|
| 1. Overview of information theory, computer science theory and stochastic thermodynamics | David H. Wolpert | The three-term split of a computation’s entropy cost |
| 7. How and what does a biological system compute? | Sonja Prohaska, Peter Stadler, Manfred Laubichler | Computation as a defining feature of living systems at every scale |
| 9. Beyond number of bit erasures | Joshua Grochow, David Wolpert | The thermodynamic cost of computation in computer science theory |
| 11. Trade-offs between cost and precision | Hildegard Meyer-Ortmanns | Aging as error that a finite energy budget cannot fully repair |
| 12. The power of being explicit | Thomas Ouldridge, Rory Brittain, Pieter Rein ten Wolde | Copying and measurement are not free, and erasure is overemphasized |
| 13. Transforming metastable memories | Paul Riechers | Speed and modularity force dissipation that ideal computation avoids |
The other chapters cover chemical computers, protocells, signal transfer between cell generations, DNA strand displacement, energy-saving software, work extraction and fluctuation theorems.
Beyond the Landauer bound
Rolf Landauer argued in 1961 that erasing one bit releases at least kT ln 2 of heat, where k is Boltzmann’s constant and T the temperature. Grochow and Wolpert restate that result in chapter 9 and then move past it. Counting erased bits, their title says, is only the start.
Wolpert’s chapter 1 gives the replacement. The entropy that a device sends into its environment while it computes splits into three terms, a decomposition he developed with Artemy Kolchinsky. The Landauer cost is the drop in Shannon entropy of the device’s states, and it depends only on the computation and the distribution of its inputs. The mismatch cost is paid when the actual inputs differ from the distribution the device was built for. The residual entropy production depends on the physical details of the device and on the structure of the computation. The second term carries the main practical point. The same device on the same task costs more energy when its inputs change, by an amount that information theory alone predicts.
Two other chapters argue against simplified versions of the field. Ouldridge, Brittain and ten Wolde reject the common belief that a measurement or a copy can be made at no cost, and they argue that erasure gets more attention than it deserves. Riechers starts from the result that computation can in theory run without dissipation and asks why real computers and brains still need fuel. His answer is that finite speed and modular design each force a cost.
What a thinking brain costs at the limit
The following calculation is this site’s, made with the framework the book sets out. At body temperature, 310 kelvin, the Landauer cost of one bit erasure is about 3.0 × 10⁻²¹ joules. A person who eats 2,000 kilocalories a day uses about 97 watts on average, so the preface’s 10 to 20 percent puts the brain at roughly 10 to 19 watts. At the Landauer limit that power would pay for about 3 × 10²¹ to 7 × 10²¹ bit erasures per second.
No one claims the brain works near that limit, and chapters 1, 12 and 13 explain why a real system cannot. Mismatch cost, residual entropy production, copying, finite speed and modularity all add to the bare Landauer term. The figure is a floor, and the gap between the floor and the brain’s real budget is the size of those other costs.
The same arithmetic applies to the hardware this site follows. The one million p-bit machine in Kerem Camsari’s FPGA computer draws 1.4 to 1.6 kilowatts for 10¹² flips per second, about 1.5 nanojoules per flip. That is roughly 500 billion times the Landauer cost of one bit at room temperature. A p-bit flip is not a bit erasure, so the ratio compares scales and no more. It shows how far present machines are from the floor that the book defines, and why companies that build chips from thermal noise claim large energy gains.
What this means for the consciousness question
This site studies consciousness as an emergent property under functionalist emergentism. That view predicts that the organization of a system carries its properties, whatever material runs it. The thermodynamics of computation adds a physical condition to that prediction. Every organization that computes pays an energy cost, and the book’s three-term account says that part of the cost depends only on the computation and part depends on the material.
The split matches the two sides of the substrate debate. The Landauer cost and the mismatch cost belong to the computation, so any substrate that runs the same organization on the same inputs must pay them. The residual entropy production belongs to the substrate. If consciousness depends only on organization, its minimum energy cost is the same in tissue and in silicon, and only the residual term differs. Todd Feinberg and Jon Mallatt argue the opposite, that real metabolic stakes are part of what makes a system sentient, as set out on the Origins of Consciousness page. On their view the cost of staying alive belongs to consciousness itself.
The book does not take a side, and it does not discuss consciousness. Its contribution is a vocabulary in which the question can be stated as a measurement. The Thermodynamic Computing hub shows the p-bit and oscillator systems that would make such a measurement possible in software, and the Kardashev scale read as an energy budget carries the same accounting up to the scale of a civilization.
Comparison to The Consciousness AI
The project measures no energy. Its p-bit code, models/thermodynamic/p_bit_emulator.py, samples an Ising model in software on ordinary hardware, and it reports no joules per sample. The cost figures in the project’s thermodynamic results count channel events, the number of p-bit samples a reading needs, which is a proxy for energy. They are not joules. The book’s framework would turn those counts into a lower bound in joules only on hardware where the residual term is known.
Strengths and limits
The volume is strongest where it is explicit. Wolpert’s decomposition gives one set of terms for cells, chemical computers and chips, and the Ouldridge chapter corrects ideas that circulate widely outside physics. The chapters are research papers for specialists across physics, computer science and biology, and a reader without stochastic thermodynamics will find chapter 1 the necessary entry point. The book appeared in 2019, so it predates the p-bit chips and FPGA machines now in development. It gives no measurement of a brain or a chip against its own bounds, and that remains open seven years later.
The Energetics of Computing in Life and Machines defines what any computing system, living or built, must pay to compute. That makes it a reference for anyone asking whether a conscious process has a minimum cost, and whether that cost depends on the material. The broader effort to agree criteria for consciousness in any system is reviewed in the race to define AI consciousness.