How thermal management is changing in the age of the kilowatt chip
Escalating chip power densities—approaching and exceeding 1 kW per package—are forcing a rethink of thermal management, from data centers constrained by per-rack power limits to enthusiasts exploring how much compute a home electrical service can realistically support. Commenters highlight emerging solutions such as advanced liquid cooling, wafer-scale processors like Cerebras’ CS-2 with extreme current delivery, and reuse of waste heat for building or district heating. While exotic high-density systems are rarely the most energy-efficient, they are seen as necessary where compute density, latency, or licensing economics outweigh pure power efficiency.
Datacenter power density and limits
- Power and cooling, not space, are seen as primary constraints. Typical commercial racks are cited around 8–10 kW historically, with newer data showing many sites moving into 16–50+ kW per rack.
- Total building power capacity and willingness of utilities/electricians to increase supply become hard limits; beyond a point it’s often cheaper to build another datacenter.
- Very large datacenters are compared to industrial loads (e.g., hundreds of megawatts), with both ultimately limited by cheap power and connectivity.
- High‑density, multi‑kW nodes make sense mostly when licensing, bandwidth, or latency benefits outweigh increased power/cooling complexity.
Residential power for high‑power compute
- Many posts map out typical household service: often 100–200 A at 230–240 V in various countries, giving tens of kilowatts total capacity.
- Individual circuits (e.g., 15–20 A outlets vs. stove/dryer or 3‑phase sockets) set practical limits for a single system.
- The 80% derating rule for continuous loads is discussed, with some disagreement on how strictly it applies in practice.
- Consensus: with upgrades, a home can support several kilowatts of compute, but cost, wiring, and space are limiting factors.
Using compute as a heat source
- Several people already heat spaces with GPUs/ML rigs or suggest doing so.
- Bitcoin mining as “smart space heaters” is debated: conceptually appealing but current mining hardware is expensive, rapidly obsoleted, and often not profitable even with very cheap electricity.
- Electric resistance heating is noted as inefficient relative to heat pumps or gas; reuse of waste heat only makes sense if the compute would run anyway.
- Examples are given of “bitcoin heaters,” a spa heated by miners, a “data furnace” research concept, and large‑scale district heating using datacenter waste heat.
Extreme cooling and the Cerebras CS‑2
- The teardown of a wafer‑scale AI system draws strong interest: ~20–23 kW into a dinner‑plate‑sized chip with roughly 850k cores.
- Discussion focuses on its 20,000 A low‑voltage power delivery, complex liquid cooling, and mechanical challenges of uniformly cooling such a large die.
- There is back‑and‑forth on realistic core voltages and currents; values around 0.7–1.2 V and tens of kA are considered plausible, though some numbers in the thread conflict or seem inconsistent.
- On‑chip DC‑DC conversion is debated; reasons against include process limitations, area overhead, inductor inefficiency, and specialized analog design expertise.
Architecture and efficiency reflections
- Some argue that ultra‑dense “kilowatt chips” are not about best compute‑per‑watt but about reducing parallel overhead and simplifying programming by keeping work on a single, very fast system.
- Ideas surface about future geometries (fractal/Sierpinski, spherical “golf ball” chips) and moves toward immersion cooling and possibly free‑space optics to reduce resistive heating in interconnects.