# EcoCompute Should-I-Quantize Estimator

> EcoCompute Should-I-Quantize Estimator is a free MCP server for ai models and llm tooling. Directional answers on whether quantizing an LLM saves GPU energy, with stated basis. It runs as a hosted remote server, so there is nothing to install locally.

- **Kind:** MCP server
- **Pricing:** Free
- **Platforms:** Web, API
- **Website:** https://quantenergy.tech
- **Repository:** https://github.com/hongping-zh/quant-energy
- **MCP Registry name:** io.github.hongping-zh/ecocompute-estimator
- **Categories:** AI models and LLM tooling
- **Listing updated:** 2026-11-15

## About

EcoCompute is an open micro-benchmark rather than a general claim about quantization. Its protocol samples GPU-package power with NVML for single-stream decoding at batch size 1 and 256 tokens, and the site says 29 of 64 planned cells have been measured. The headline finding is a crossover that depends on model size and hardware, not a rule that NF4 or INT8 always saves energy.

Give the estimator a GPU and model size and it returns a directional verdict on energy per million tokens compared with FP16. Tested configurations are shown as measured, while other sizes come from a fitted curve whose uncertainty is displayed; points measured only once are marked as single observations. INT8 was measured on A800 and RTX 4090 only.

The page asks readers to treat results as directional rather than a production decision. A container and a free Colab T4 route let you produce your own energy.json, and the dataset is archived on Zenodo under CC BY 4.0.

## Features

- **Directional verdict:** Says whether quantization saves or costs energy for a chosen GPU and model size.
- **Stated basis:** Each number is labelled as measured or interpolated.
- **Crossover curve:** Shows how savings change with model size, with uncertainty on fitted points.
- **Optimize search:** Searches precision, batch and context under your constraints.
- **Reproduce it yourself:** A runnable container and Colab path produce your own measurement.
- **Citable dataset:** Archived dataset with DOI, protocol and limitations statement.

## Good for

- Decide whether to run a 7B model in NF4 on a given GPU
- Check how far a result generalises before citing it
- Reproduce an energy measurement on your own card
- Cite benchmark data in a paper on inference energy

## Install

### Claude Code (Terminal)

```sh
claude mcp add --transport http ecocompute-estimator https://ecocompute-estimator.zhanghongping1982.workers.dev/mcp
```

### Claude Desktop (Settings → Connectors → Add custom connector)

```sh
https://ecocompute-estimator.zhanghongping1982.workers.dev/mcp
```

### Cursor (~/.cursor/mcp.json)

```json
{
  "mcpServers": {
    "ecocompute-estimator": {
      "url": "https://ecocompute-estimator.zhanghongping1982.workers.dev/mcp"
    }
  }
}
```

### VS Code (.vscode/mcp.json)

```json
{
  "servers": {
    "ecocompute-estimator": {
      "type": "http",
      "url": "https://ecocompute-estimator.zhanghongping1982.workers.dev/mcp"
    }
  }
}
```

### Windsurf (~/.codeium/windsurf/mcp_config.json)

```json
{
  "mcpServers": {
    "ecocompute-estimator": {
      "serverUrl": "https://ecocompute-estimator.zhanghongping1982.workers.dev/mcp"
    }
  }
}
```

### Other clients (your client's MCP settings (most use this shape))

```json
{
  "mcpServers": {
    "ecocompute-estimator": {
      "type": "http",
      "url": "https://ecocompute-estimator.zhanghongping1982.workers.dev/mcp"
    }
  }
}
```

## FAQ

### Is this a universal law?

No. The site limits claims to the GPUs, stacks, batch sizes and contexts measured so far.

### How is power measured?

By NVML GPU-package power sampling, not TDP estimates or wall-plug readings.

### Can I verify it without a GPU?

Yes, a free Colab T4 can produce an energy.json.

### Where is the data?

On Zenodo under CC BY 4.0.

### Is EcoCompute Should-I-Quantize Estimator free?

Yes. EcoCompute Should-I-Quantize Estimator is free to use.

### How do I add EcoCompute Should-I-Quantize Estimator to Claude Code?

Run this command in your terminal: claude mcp add --transport http ecocompute-estimator https://ecocompute-estimator.zhanghongping1982.workers.dev/mcp

### What are the best alternatives to EcoCompute Should-I-Quantize Estimator?

The closest alternatives to EcoCompute Should-I-Quantize Estimator are Claude Code Ultimate Guide, signafAIde, Toolkien, Lugh and HypePaper.

## Alternatives

- [Claude Code Ultimate Guide](https://ai-powered.com/mcp/claude-code-ultimate-guide): Claude Code Ultimate Guide is an open-source MCP server for ai models and llm tooling.
- [signafAIde](https://ai-powered.com/mcp/signafaide): signafAIde is a freemium MCP server for security.
- [Toolkien](https://ai-powered.com/mcp/toolkien): Toolkien is an open-source MCP server for ai models and llm tooling.
- [Lugh](https://ai-powered.com/mcp/lugh): Lugh is an open-source MCP server for ai models and llm tooling.
- [HypePaper](https://ai-powered.com/mcp/hypepaper): HypePaper is a freemium MCP server for ai models and llm tooling.
- [Agent Data Tools: Compact (find, describe, run)](https://ai-powered.com/mcp/agent-data-tools-compact-find-describe-run): Agent Data Tools: Compact (find, describe, run) is a freemium MCP server for developer tools.
- [Angel Agent Guard: pre-purchase check + prompt-injection scan](https://ai-powered.com/mcp/angel-agent-guard-pre-purchase-check-prompt-injection-scan): Angel Agent Guard: pre-purchase check + prompt-injection scan is a freemium MCP server for ai models and llm tooling.
- [Mizan Calibration](https://ai-powered.com/mcp/mizan-calibration): Mizan Calibration is a free MCP server for ai models and llm tooling.

---
Source: https://ai-powered.com/mcp/ecocompute-should-i-quantize-estimator
