Φ

VOLTAGE → BYTECODE

The Engineering

NUMEN is an integer-only computing substrate. Eight layers, from the transistor to the bytecode, with no floating-point unit, no GPU, and no math library. What follows is what it does, where the speed comes from, and how to reproduce every number.

no FPUno GPUno libmno backpropbit-exactO(1) recall

What it actually does

It stores, recalls, learns, schedules, and renders — using nothing but 64-bit integers. Every value lives in Q32.32 fixed point. Every address is resolved by a golden-ratio hash in constant time. Every learning step is a contraction that is proven to converge. The result is a system that produces the exact same bytes on a laptop, a microcontroller, or a server.

There is no hidden accelerator doing the heavy lifting. The speed is structural: work is either O(1) or eliminated before it runs.

Grounded recall
Constant-time key resolution — never a scan.
Gradient-free learning
Banach contraction, guaranteed to converge.
Prime scheduling
Fires only when the clock hits a prime.
Integer graphics
A full renderer with no GPU and no OpenGL.

Core measured results

Six headline numbers. Each was measured on real hardware and is reproducible from source.

0.414 J
Energy to train
One full learning run on a $30 board. ~450M× below a transformer.
5,759×
Grounded recall speedup
16.33 ns grounded vs. 94,031 ns ungrounded.
2.32 Gsteps/s
Edge throughput
Sustained on a Teensy 4.1 at 600 MHz.
112.95×
Scheduler efficiency
99.11% of pathways eliminated before they run.
1,050 FPS
Integer graphics
924.7M voxels/s on one CPU core. No GPU.
bit-exact
Determinism
Identical Q32.32 output on every platform, SHA-256 sealed.

O(1) grounded recall — where the speed lives

The single most important primitive is the lookup. NUMEN resolves any key through a multiplicative golden-ratio hash that produces zero collisions across the entire address space. There is no probing, no chaining, no scan — the cost of finding an entry is the same whether the table holds ten items or a hundred thousand.

16.33 ns
per grounded lookup
94,031 ns
ungrounded equivalent
0.000%
collision rate
200,000 / 200,000
exact hits

Why it is genuinely O(1)

The hash multiplies the key by the 64-bit golden-ratio constant 0x19E3779B97F4A7C1 and takes the high bits. Because the multiplier is coprime to the table size and spreads keys by the golden angle, every slot is used and no two keys ever land together. The full key-resolution chain — hash, index, verify — was measured at 25.7 ns end to end.

What 5,759× means in practice

Grounded recall runs at 16.33 ns; the ungrounded path that has to search costs 94,031 ns. That is a 5,759× gap on the same machine, over 146,463 live entries. The speedup is not from a faster clock — it is from never doing the search at all.

The eight-layer stack

From the voltage on the silicon up to the symbolic bytecode. Each layer does one job and hands a deterministic result to the next.

0.1
Voltage & Silicon (MSR Substrate)

Maps transistor switching straight into Q32.32 fixed point. No floating-point unit is ever touched — the substrate is defined at the level of the silicon.

PHI_FP = 0x19E3779B97F4A7C1SOURCE/core/phi_fixed.h
0.2
Prime Addressing

Collision-free golden-ratio hashing. True O(1) key resolution — the lookup cost does not grow with the table.

16.3 ns / lookup over 146,463 entries · 0.000% collisionsSOURCE/core/rel_address.h
0.3
Phi-Net Convergence

Eight independent poles — FAST, EXACT, VAST, EMBODIED, RENDER, NAVIGATION, IDENTITY, SYSTEM — all converge to the same constant.

φ⁻¹ = 0.618033SOURCE/core/phi_net.h
0.4
Q32.32 Arithmetic Floor

64-bit fixed point, 32 integer / 32 fractional bits. Zero float anywhere in the pipeline. Deterministic multiply, divide, and Newton square root.

range ±2,147,483,647 · ~0.23 nanounit precisionSOURCE/core/fp_mul · fp_div · fp_sqrt_newton
0.5
CORDIC Trigonometry

28-iteration shift-and-add trig. No libm, no lookup tables — sine, cosine, and rotation built from integer shifts.

112.95× faster than libm on x86-64SOURCE/core/cordic.h
0.6
Banach Learning (gradient-free)

Fixed-point contraction with Lipschitz constant k < 1. Convergence is mathematically guaranteed — no gradients, no backprop, integer-only.

0.414 J to full convergence (Teensy 4.1 @ 600 MHz)SOURCE/quatos/quatos_l3_trainer.c
0.7
Prime-Rhythm Scheduler

256 channels, one prime each. A channel fires only when t mod Pᵢ == 0, so most work is eliminated before it ever executes.

2.266444 firings/tick measured vs. Σ(1/Pᵢ) = 2.266451 analytic · 99.11% eliminatedSOURCE/experiments/metal_is_os/the_metal_is_the_os.c
0.8
ReL-256 Bytecode

A 256-opcode symbolic layer. Every opcode maps directly to one Zone-0 operation — the language and the metal are the same thing.

256 opcodes · 1:1 to Zone-0SOURCE/core/rel_axiom.h · rel_sigma256.h

Prime-Rhythm Scheduler

256 channels, one prime per channel. A channel only fires when the tick count is divisible by its prime, so 99.11% of the potential work is eliminated before it runs. Over 20M ticks the scheduler produced 45,328,885 firings — a measured 2.266444 firings/tick against the analytic Σ(1/Pᵢ) = 2.266451. The two agree to six figures without a single number being hand-tuned.

99.11%
pathways eliminated
112.95×
efficiency gain
45,328,885
firings / 20M ticks
≈ 10⁶⁸⁹
grand downbeat

Gradient-free learning

Learning is a Banach fixed-point contraction with Lipschitz constant k < 1. Because the map is a contraction, it is mathematically guaranteed to converge to a unique fixed point — no gradients, no backpropagation, no floating point. A full training run reaches convergence on a $30 microcontroller using 0.414 J of energy, roughly 450 million times less than training a comparable transformer.

0.414 J
per training run
k < 1
Lipschitz constant
~450M×
below a transformer
integer
no float, no backprop

Reproducibility & provenance

Every number on this page is tagged at the source as [M]easured, [P]roven, or [I]nterpretation. No number is hand-typed to make two results agree. The full codebase compiles with a stock C compiler — no framework, no runtime, no dependencies.

14,348
lines of source
31
modules
5
proof suites
8
corpora
reproduce the headline benchmarks
cd SOURCE/experiments/disc_demonstration
cc -O2 speedup_bench.c -o speedup_bench && ./speedup_bench
 
cd SOURCE/experiments/metal_is_os
cc -O2 the_metal_is_the_os.c -lm -o the_metal_is_the_os
./the_metal_is_the_os

Published across 23 records on Zenodo · 491 downloads · CC BY-NC-ND 4.0 · ORCID 0009-0003-5519-7939. Every run is SHA-256 sealed for bit-exact verification.

See it run

The experiments page runs the O(1) hash and φ-convergence live in your browser. The gallery shows the bare-metal renders and the architecture behind them.