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.