EcoCompute Should-I-Quantize Estimator
Energy impact of LLM quantization, calibrated to measured GPU data. Every answer states its basis.
我該用這個嗎
品質與安全性
根據工具定義與協定合規性的自動化分析。
上下文成本
這是每次將伺服器的工具載入模型上下文時所消耗的約略 token 數量。數量越高，可用於其他工作的注意力就越少。
安裝
一鍵安裝
將以下內容加入你的 `claude_desktop_config.json` 檔案：
{
"mcpServers": {
"ecocompute-estimator": {
"url": "https://ecocompute-estimator.zhanghongping1982.workers.dev/mcp"
}
}
}遠端端點
https://ecocompute-estimator.zhanghongping1982.workers.dev/mcpstreamable-http它能做什麼
工具清單
工具（4）
🟡estimate_quantization_energy(params_b, arch, precision, batch, ctx, ...)
Estimate the task-level energy change (ΔE% vs FP16) of weight-only quantization (NF4 or INT8) for a model of a given size on a GPU architecture class, from measured-anchored curves. The answer carries basis (measured/interpolated/estimated), confidence and evidence boundaries.
輸入結構描述
{
"type": "object",
"properties": {
"params_b": {
"type": "number",
"description": "model size in billions of parameters, e.g. 7"
},
"arch": {
"type": "string",
"enum": [
"turing",
"ada",
"blackwell",
"ampere",
"hopper"
],
"description": "GPU architecture class"
},
"precision": {
"type": "string",
"enum": [
"NF4",
"INT8"
],
"description": "quantization format (default NF4)"
},
"batch": {
"type": "integer",
"description": "batch size (default 1; >1 shrinks savings — modelled, not measured)"
},
"ctx": {
"type": "integer",
"description": "context length (default 2048)"
},
"q": {
"type": "string",
"description": "alternative: plain-English query, e.g. '13B on A100 int8 batch 32'"
}
}
}🟢get_quantization_crossover(arch)
Get the model-size crossover for NF4/INT8 per architecture: below it quantization costs energy, above it saves. Crossovers belong to a card + software stack, not an architecture — the answer says so.
輸入結構描述
{
"type": "object",
"properties": {
"arch": {
"type": "string",
"enum": [
"turing",
"ada",
"blackwell",
"ampere",
"hopper"
],
"description": "optional: limit to one architecture"
}
}
}🟢list_data_coverage
List what is actually measured (GPU architecture × precision × size range, NVML direct), what is borrowed/estimated, and where the gaps are. Use this before trusting an estimate at the edges of coverage, and to answer 'is there data for my GPU?' honestly.
輸入結構描述
{
"type": "object",
"properties": {}
}🟢search_measurements(query)
Search the measured anchors by model name, GPU name, architecture or size. Returns ΔE% vs FP16 per anchor with run counts — the raw evidence behind the curves.
輸入結構描述
{
"type": "object",
"properties": {
"query": {
"type": "string",
"description": "model ('TinyLlama'), GPU ('4090'), architecture ('ampere') or size ('7')"
}
},
"required": [
"query"
]
}社群
證據