green-roof — cost / carbon / SLO calculator

Config

Model GPU (built-in $/hr · BW · W)

Batch 8 (1–64) Seq len 2048 (256–8192) SLO p99 budget 250 ms (50–1000) QPS target 10 (1–100)

Outputs — selected GPU

Throughput

–

$/1M tokens

–

Energy / 1M

–

gCO₂e / 1M

–

Replicas for SLO

–

Feasible

–

Tradeoff — $ vs CO₂ across GPUs (log–log)

GPU	tok/s	$/1M	gCO₂/1M	p99 est	replicas	SLO

Green = SLO-feasible, red × = miss. Star ★ = selected. Hover title shows $/hr · BW · W.

Specs from green_roof/gpu.py + model.py + planner.py (bf16 weights, util 0.6/0.7, PUE 1.15, grid 380 g/kWh, 128 tok/req, p99≈2×mean). deepseek-16b (16B, KV 640 B/tok) is a back-of-envelope estimate, same method as MODEL_ZOO.