Environment & CO2 methodology

SuperCompress shortens prompts before inference. It does not read or write model KV cache. Fewer input tokens means less prefill work on the GPU for the same workflow, lower energy and CO₂ as a side effect of sending less text.

These numbers are illustrative estimates, not per-deployment measurements. All assumptions are explicitly documented so you can adjust them for your hardware and energy mix.

What we measure

Metric	Definition
Tokens saved	original_tokens - kept_tokens per compression
Token savings % (tokens_saved_pct)	(1 - kept/original) * 100, percent of prompt tokens removed (historical field name; not an in-KV operation)
GPU-seconds avoided	Effective tokens saved / throughput
Wh saved	GPU-seconds * GPU watts / 3600
CO2 avoided	Wh * grid intensity (kg/kWh) / 1000
Water saved (est.)	Datacenter cooling water estimate (evaporative cooling)

Default assumptions

Defined in supercompress/benchmarks/metrics.py:

Parameter	Default	Rationale
tokens_per_gpu_second	2,500	7B-class model prefill on consumer GPU (e.g., RTX 4090)
gpu_watts	150 W	Typical single-GPU draw during inference (not peak)
context_share_of_prefill	55%	Attribution share of prefill cost treated as context-driven when estimating energy, SuperCompress does not edit KV cache
grid_kg_co2_per_kwh	0.417	US grid average (EIA 2024 data)

Python sustainability API

The sustainability_from_tokens_saved helper converts token savings into estimated energy and carbon metrics.

from supercompress.benchmarks.metrics import sustainability_from_tokens_saved

saved = result.original_tokens - result.kept_tokens
impact = sustainability_from_tokens_saved(saved)

print(impact.watt_hours_saved)
print(impact.co2_kg_avoided)
print(impact.water_liters_saved)
print(impact.assumptions.to_dict())

The impact.assumptions object exposes the four constants above so you can modify them for your specific hardware and region.

Scale examples

Based on our default assumptions, at ~800 tokens saved per compression:

Scale	Tokens avoided	kWh saved	CO2 avoided	Water saved (est.)
1 compression	~800	~0.00003	~0.00001 kg	~0.0001 L
1K compressions	~800K	~0.03	~0.01 kg	~0.1 L
1M compressions	~800M	~29	~12 kg	~100 L
10M compressions	~8B	~290	~120 kg	~1,000 L
100M compressions	~80B	~2,900	~1,200 kg	~10,000 L

Why CPU eviction matters

The learned policy (a learned query-aware neural policy) runs on CPU before GPU inference. This means:

• The eviction itself consumes negligible energy, ~0.000001 kWh per compression on a modern CPU core
• But it avoids much larger GPU prefill work on long contexts
• Fewer prompt tokens mean less GPU prefill: CPU-side selection is cheap compared with the inference it avoids

This 1:1000 energy ratio is why CPU-side eviction is environmentally significant. The tiny policy makes it possible, a 165M-parameter model running on CPU would consume much more energy per eviction, eroding the net savings.

Honesty for submissions

We follow these principles when presenting environmental impact data:

1. State assumptions clearly, never claim live datacenter metering unless you have it
2. Compare quality + savings together, truncation can save tokens but drops the middle regardless of the question, while SuperCompress picks lines against the question
3. Don't claim CO2 without methodology, every estimate links back to the documented assumptions and formulas
4. CPU energy is negligible but included, hosted Neural Keep runs on CPU; the energy cost is orders of magnitude below the GPU prefill savings
5. Let users adjust constants, the Python API exposes all assumptions so users can plug in their own numbers

See benchmarks page for the interactive projection calculator and supercompress/benchmarks/metrics.py for the source code.