Energy impact of LLM quantization, calibrated to measured GPU data. Every answer states its basis.
Maintain this server? Claim it
Call this server's tools from your browser. Nothing to install.
Requests go through mcptry because browsers can't call most MCP servers directly. Header values are sent to the server for that request only and are never stored or logged.
Tool list from our last check. Connect live to refresh it.
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.
alternative: plain-English query, e.g. '13B on A100 int8 batch 32'
context length (default 2048)
GPU architecture class
batch size (default 1; >1 shrinks savings — modelled, not measured)
model size in billions of parameters, e.g. 7
quantization format (default NF4)
This tool isn't marked read-only, so it may change something on the server or in an account you connected.
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.
| Argument | Type | Description |
|---|---|---|
| q | string | alternative: plain-English query, e.g. '13B on A100 int8 batch 32' |
| ctx | integer | context length (default 2048) |
| arch | string | GPU architecture class |
| batch | integer | batch size (default 1; >1 shrinks savings — modelled, not measured) |
| params_b | number | model size in billions of parameters, e.g. 7 |
| precision | string | quantization format (default NF4) |
{
"type": "object",
"properties": {
"q": {
"type": "string",
"description": "alternative: plain-English query, e.g. '13B on A100 int8 batch 32'"
},
"ctx": {
"type": "integer",
"description": "context length (default 2048)"
},
"arch": {
"enum": [
"turing",
"ada",
"blackwell",
"ampere",
"hopper"
],
"type": "string",
"description": "GPU architecture class"
},
"batch": {
"type": "integer",
"description": "batch size (default 1; >1 shrinks savings — modelled, not measured)"
},
"params_b": {
"type": "number",
"description": "model size in billions of parameters, e.g. 7"
},
"precision": {
"enum": [
"NF4",
"INT8"
],
"type": "string",
"description": "quantization format (default NF4)"
}
}
}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.
| Argument | Type | Description |
|---|---|---|
| arch | string | optional: limit to one architecture |
{
"type": "object",
"properties": {
"arch": {
"enum": [
"turing",
"ada",
"blackwell",
"ampere",
"hopper"
],
"type": "string",
"description": "optional: limit to one architecture"
}
}
}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.
No arguments.
{
"type": "object",
"properties": {}
}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.
| Argument | Type | Description |
|---|---|---|
| query* | string | model ('TinyLlama'), GPU ('4090'), architecture ('ampere') or size ('7') |
{
"type": "object",
"required": [
"query"
],
"properties": {
"query": {
"type": "string",
"description": "model ('TinyLlama'), GPU ('4090'), architecture ('ampere') or size ('7')"
}
}
}