A New Way to Match Computers to Physical Hardware

In the paper's words

Computational architectures are today chosen for a substrate by benchmarking, an expensive process that must be repeated for every substrate-architecture pair and gives no guidance before hardware exists. No existing framework predicts, from thermodynamics alone, which architecture classes a physical substrate can sustain. We propose the Causal-Energy Invariant (CEI): for a given substrate, the product of causal density ρ_t (dependent operations per volume-time) and energy fidelity η_e (energy per transition in units of kTln 2) is bounded by a substrate constant κ, derived from Landauer's principle, a substrate power-density limit, and causal-cone bounds on operation rates.

The abstract, verbatim, and its opening sentence is the claim. Every line below is quoted from the text.

Numbers it reports

1. We derive $\kappa = 3.483\times 10^{29}\ \mathrm{m^{-3}\,s^{-1}}$ for silicon under an assumed areal power limit of $100\ \mathrm{W\,cm^{-2}}$ and a $1\ \mathrm{mm}$ active depth, show the causal-cone bound exceeds this by a factor $\approx 8.6\times 10^{10}$ (so power, not relativity, binds silicon), and classify two assumed architecture profiles: a pipelined CMOS-like profile saturating $3.0\%$ of $\kappa$ and a gate-based error-corrected profile with $\eta_e \geq 100$ from published thermodynamic floors.

   from Abstract

2. The workloads they care most about — combinatorial optimization, privacy-preserving computation, large-scale inference — are dominated by hardware whose cost scales with the number of discrete operations executed, and special-purpose accelerators, neuromorphic chips, and optical processors are today evaluated by qubit or transistor counts rather than by any physical measure of fit between machine and medium [13].

   from 1. Introduction

3. Energy-per-operation measures and operation-rate measures exist separately, and Landauer's principle gives a floor for the former [13], but no published framework composes them into a single invariant that partitions architectures into equivalence classes per substrate.

   from 1. Introduction

4. A worked numerical classification for silicon under explicitly stated assumptions, giving $\kappa = 3.483\times 10^{29}\ \mathrm{m^{-3}\,s^{-1}}$, a causal-to-power bound ratio of $8.607\times 10^{10}$, and two architecture placements at $3.0\%$ of $\kappa$ and, in principle, at the bound itself ($\Pi_G/\kappa \leq 1$) (Section 5).

   from 1. Introduction

5. Input: Boltzmann constant $k = 1.380649\times 10^{-23}\ \mathrm{J\,K^{-1}}$; $T = 300\ \mathrm{K}$ (Assumption A4); $\ln 2 = 0.6931472$. $$kT \;=\; (1.380649\times 10^{-23})\times 300 \;=\; 4.141947\times 10^{-21}\ \mathrm{J},$$ $$kT\ln 2 \;=\; (4.141947\times 10^{-21})\times 0.6931472 \;=\; 2.871\times 10^{-21}\ \mathrm{J}.$$ This reproduces the single-bit erasure floor derived in [13], which states the same value $2.871\times 10^{-21}\ \mathrm{J}$ at $300\ \mathrm{K}$, computed from $kT\ln 2$.

   from 4.1 The Landauer constant

6. By Assumption A3, a dependent operation occupies at least a volume $L^3$ and at least a time $L/c$, so $$\rho_t^{\mathrm{causal}} \;=\; \frac{c}{L^4}.$$ With $c = 2.998\times 10^{8}\ \mathrm{m\,s^{-1}}$ and $L = 10\ \mathrm{nm} = 1.0\times 10^{-8}\ \mathrm{m}$ (Assumption A4): $$L^4 = (1.0\times 10^{-8})^4 = 1.0\times 10^{-32}\ \mathrm{m^4},$$ $$\rho_t^{\mathrm{causal}} = \frac{2.998\times 10^{8}}{1.0\times 10^{-32}} = 2.998\times 10^{40}\ \mathrm{m^{-3}\,s^{-1}}.$$ The ratio of the causal bound to the power bound: $$\frac{\rho_t^{\mathrm{causal}}}{\kappa} \;=\; \frac{2.998\times 10^{40}}{3.483\times 10^{29}} \;=\; 8.607\times 10^{10},$$ where $\kappa$ is computed in Section 4.5 below.

   from 4.3 The causal-cone bound and which constraint binds

How it is done

1. Architectures with equal $\rho_t\eta_e$ form genuine equivalence classes, and the bound predicts that substrates favor architectures whose product approaches $\kappa$.

   from Abstract

2. The contributions of this paper are: A derivation of the CEI bound $\kappa$ from Landauer's principle, a substrate power-density limit, and causal-cone bounds on operation rates, with all arithmetic shown (Section 4).

   from 1. Introduction

3. The formalization of quotients in [4] treats the computational content of equivalence classes: defining a function on a quotient requires proving the result independent of the chosen representative, a proof obligation that users of theorem provers discharge by preferring canonical representatives such as de Bruijn variables for $\alpha$-equivalence.

   from 2. Background and Related Work

4. This paper is analytical: it defines a model of substrate-architecture interaction, derives the CEI bound from three physical inputs, proves the associated relation is an equivalence relation, and computes numerical classifications under explicitly stated assumptions.

   from 3. Methods

What would falsify it

1. No existing framework predicts, from thermodynamics alone, which architecture classes a physical substrate can sustain.

   from Abstract

2. Architectures with equal $\rho_t\eta_e$ form genuine equivalence classes, and the bound predicts that substrates favor architectures whose product approaches $\kappa$.

   from Abstract

3. Confirmation would replace per-pair benchmarking with a thermodynamic substrate-architecture map; falsification would localize where the tradeoff breaks.

   from Abstract

Where the claim stops

1. What the invariant does and does not claim.

   from 6. Discussion

2. It does not claim that all architectures on a substrate converge to $\kappa$, nor that a high $\Pi$ is desirable independent of workload — a profile can saturate the bound while solving the wrong problem.

   from 6. Discussion

3. CEI is deliberately single-scalar, and this is its principal limitation: two architectures in the same $\Pi$ class may differ arbitrarily in latency structure, fault behavior, and programmability.

   from 6. Discussion

Where to read

1. Abstract202 words · 1 min
2. 1. Introduction614 words · 3 min
3. 2. Background and Related Work1,042 words · 5 min
4. 3. Methods387 words · 2 min
5. 4. Analysis1,162 words · 5 min
6. 5. Results299 words · 1 min
7. 6. Discussion780 words · 3 min
8. 7. Practical Implications and Engineering Pathways607 words · 3 min
9. 8. Conclusion222 words · 1 min
10. Provenance and AI Involvement225 words · 1 min