lablab.ai Hackathon 2026AMD

GreenTune Agent

GeminiMI300X

Autonomous Energy Intelligence for LLM fine-tuning — powered by Gemini on AMD MI300X

Multi-agent swarm optimizes hyperparameters to minimize Joules-per-token, enforcing energy governance policies in real-time. Built with real AMD MI300X telemetry data from QLoRA fine-tuning runs.

Key FindingAMD MI300X

30% More Energy Wasted

Smaller batch sizes do not reduce power draw on MI300X. The GPU saturates at ~750W regardless — smaller batches just take longer, burning 26,534 extra Joules for the same result.

Baseline (bs=2)

87,300

Joules total

Small Batch (bs=1)

113,834

Joules total

J/Token (Baseline)

0.355

Joules per token

J/Token (Small)

0.463

Joules per token

GPU Power Draw — Both Runs Overlaid

AMD MI300X · 750W TDP

Same power draw (~750W peak) but the small-batch run runs 40s longer.The red zone is wasted energy.

Cumulative Energy — Where the Waste Happens

Both runs process the same 245K tokens. The gap between the curves is pure waste.

AMD

Instinct MI300X

CDNA3 Architecture

VRAM

192 GB HBM3

TDP

750W

ROCm

7.0

Arch

gfx942

Carbon Impact

Baseline run

9.5g CO2

1.2 smartphone charges

Energy waste per run

+2.9g CO2

0.0071 car-miles equivalent

At Scale (1K runs/month)

Energy saved

7.4 kWh/mo

26,534 J x 1,000 runs

CO2 avoided

2.9 kg CO2/mo

7 car-miles avoided

Methodology: Power sampled via amdsmi.amdsmi_get_power_info() every 0.5s. Energy calculated using trapezoidal integration. CO2 factor: 390 gCO2/kWh (US average). Both runs: Qwen2.5-7B, QLoRA NF4, LoRA r=16, 500 Hermes traces, effective batch size 8.

Baseline (bs=2, ga=4)

Small Batch (bs=1, ga=8)