Siddha

Sustainability

Frugality is the method,
not the marketing.

We don't offset our compute. We mostly don't use it. A model that synthesizes in about a minute on one CPU core makes the energy question nearly disappear — and that is a property of the mathematics, not a pledge.

One model build

The arithmetic, honestly stated.

For a comparable small classification workload, full RSN synthesis measured ≈60 seconds on a CPU (~35 W) — about 0.0006 kWh. A modest 2-hour fine-tune on one A100 (~400 W) is ≈0.8 kWh — three orders of magnitude more. We state the comparison narrowly: it holds for the discriminative workloads where synthesis competes, and we make no energy claims about workloads it cannot do.

Energy per comparable model build
Siddha synthesis · 60 s CPU≈ 0.0006 kWh
Fine-tune · 2 h, 1× A100≈ 0.8 kWh
Training run · 8 h, 4× A100≈ 13 kWh

Synthesis time measured in the RSN benchmarks; GPU figures are illustrative magnitudes from public hardware power ratings. The honest comparison class is small discriminative models — not frontier LLMs, which synthesis cannot replace.

Why the footprint is structural
01

No training loop

The energy of deep learning is dominated by optimization — thousands of gradient passes over the data on power-hungry accelerators. Synthesis makes one statistical pass on a CPU. The loop, and its megawatts, simply never exist.

02

No specialized hardware

Every RSN result we publish was produced on an ordinary consumer CPU — the research machine had 8 GB of RAM. No accelerator was manufactured, shipped, cooled, or retired for any of it.

03

No retraining debt

Because a rebuild costs seconds, models can follow drifting data without the standing energy cost of scheduled re-training pipelines. The greenest GPU-hour is the one never booked.

04

Models that run where data lives

Synthesis is light enough for the edge — a browser tab, a clinic laptop, a field device. No data shipped to a datacenter also means no datacenter on the bill, energetic or otherwise.

The Siddha view

Nature does not train its intelligence by brute force. A leaf computes its venation from local rules, in one growing season, on sunlight. We think at least some of machine intelligence can be grown the same way — and where it can't, we'll publish that too.