Useful Load => Useful Heat · Basel 2025 Thermal-Compute 3D Atlas
Can 4 x 1 kW liquid-cooled servers heat a Basel home while doing real Digital Humanities compute?
Basel Klingental 2025
swisstopo + SpeedTree · 31.2 MWh/yr demand
Liquid-Cooled Plant Room
4 x 1 kW rack · Liebert CDU · 35 kWh buffer
Gartenbad St. Jakob
Summer sink archetype · 96.2% utilisation
Vaibhav Motwani · Prakash Aryan · Vilas Harapanahalli
Live looping videos in HTML · press Space / Arrow keys · press N for speaker notes
1 / 8Open with CalorGrid: three real ray-traced scenes of our Basel reference neighborhood, liquid-cooled plant room and summer Verbund pool sink. Every number in this talk comes from one verified 8,760-hour Python conservation engine.
4 x 1 kW servers · 35 kWh lossless tank · IWB H4 0.3325 CHF/kWh excl. VAT · scored DH-Lab trace (12 srv-h/srv/day)
80% recovered to hydronic loop (3.2 kW_th) · 20% to 1-node lumped basement air · no heat pump
Sources: Open-Meteo Basel 2025 UTC->Europe/Zurich (8,760 h, 206 heating days, -5.6 C min, 31,231 kWh); ElCom 2026 IWB H4 (0.3325 CHF/kWh excl. VAT; displaced heat value 0.133 CHF/kWh_th).
All 7 required reference-house numbers upfront: 4x1 kW servers, 34.1% heat covered, 0.0 kWh dumped vs 3,790 kWh baseline, 38.0% util, 26.3% missed deadlines, CHF 4,424 elec vs CHF 3,326 income and 0.2261 CHF/srv-h OPEX break-even.
The exact same Python function (decide_live_step()) drives the 8,760 h balance and real CPU Digital Humanities jobs
tesseract PID; heat demand resumes the same PID.
80% recovered server heat feeds back into the 35 kWh buffer tank (closing the hourly conservation loop)
Verified day runs: Jan 15 (212.59 kWh demand, 38.40 kWh used, 0.0 kWh dumped); Jul 15 Smart (15.0 kWh elec, 12.0 kWh used, 0.0 kWh dumped) vs Baseline (48.0 kWh elec, 26.4 kWh dumped).
All three videos loop live right inside the slide: winter peak cutaway, summer dump bypass and our live local CPU tesseract process pausing on SIGSTOP and resuming on SIGCONT.
Smart scheduling stops summer heat dumping, but 3,816 warm hours leave expensive server hardware idle
Continuous Annual Heat Dumped (N=4)
Heat-Led Server Utilisation (Unlimited Queue)
Comparator note: Unlimited-queue runs isolate each thermal sink's absorption ceiling (64.6% single-house util vs 38.0% on the arrival-limited Slide-2 DH-Lab trace).
This visual explains the central bottleneck: pausing servers solves dumping, not hardware amortisation. Adding a summer sink lifts utilisation from 64.6% to 96.2%-100%.
3-way diverter sends July surplus to borehole recharge or Swiss warm-water sites (central CAPEX 1,400 CHF/srv-yr)
| Archetype (N=4, unlimited queue) | Loop | W45 | Util | CHF/srv-h |
|---|---|---|---|---|
| Basel Single House (comparator) | 55/40 C | No | 64.6% | 0.473 |
| Dual-Sink Erdsonde (SIA 384/6) | 40/30 C | Sec | 96.2% | 0.392-0.43* |
| Paired Freibad (St. Jakob) | 32/26 C | Yes | 96.2% | 0.392-0.43* |
| Sole Uno Bath / ARA Biogas (24/7) | 42-45 C | Yes | 100.0% | 0.386 |
| Davos Alpine House (291 heat days) | 55/40 C | No | 84.1% | 0.416 |
All 7 Verbund scenes close hourly conservation (residual <= 7.3e-9 kWh). Laundry (60/45 C) and brewery CIP (62/47 C) are preheat-only outside W45; grid feed-in stays out of scope.
Explain why pairing one house with an underground Erdsonde borehole or placing servers at year-round 35-45 C warm-water sites lifts utilisation to 96%-100% and brings total break-even down to 0.386-0.43 CHF/srv-h.
Five measured limits kept outside the lossless benchmark
1. HEATWISE Idle-Power Reversal (21.0-26.8% idle)
Long pauses need deep sleep or power-off, not OS SIGSTOP.
2. ElCom 2026 Retail H4 vs Industrial C6 Tariff
H4 gap erases 75% of the 0.1064 heat credit; COP 3.5 lifts OPEX floor to 0.2565.
HEATWISE (Zenodo 23237930): hybrid liquid 54.9 C vs 91.2 C air (9.4% slowdown); 80% capture at 50-60 C needs Aquasar full-board cooling. MLPerf v5.1: derived energy-only bounds (not verified by MLCommons).
Covers five hard engineering limits: paused idle power reversal, retail vs industrial tariffs, N=12 CAPEX cliff, ASHRAE W45 temperature limits and MLPerf v5.1 energy-only marginal cost.
One FastAPI contract · zero browser math · 4-color 3D overlay labelling each mesh by source
SIGSTOP/SIGCONT transitions.
Attributions: (c) swisstopo; (c) OpenStreetMap contributors, ODbL; (C) buildingSMART International Ltd. (CC BY 4.0); NVIDIA, Vertiv and ForgeCAD SimReady (CC BY 4.0); Three.js (MIT); Tesseract (Apache-2.0).
Provenance Mode answers visual skepticism in one click: green shows live OS worker telemetry, amber marks benchmark inputs, blue marks Swiss geodata and open CAD stand-ins inside an 8.5 MB CPU clone.
Basel pilot targets from 8,760 h physics, measured idle power and ElCom tariffs
uv sync && uv run digitalheat verify && uv run digitalheat serveLiterature: ETH Zurich Aquasar (Zimmermann et al., Energy 2012); Data Furnace (Liu et al., USENIX HotCloud 2011). Datasets: Open-Meteo, ElCom/LINDAS 2026, HEATWISE, MLPerf v5.1.
Close on two first pilot targets: year-round 35-45 C warm-water hosts and same-server Erdsonde borehole homes, gated by the 3-point pre-hardware checklist.