# Cold Plate A cold plate is a block of metal with liquid flowing through it, clamped onto a chip. Everything sophisticated about liquid cooling happens in the few millimeters between the silicon and the coolant, and the decisive part of that distance is a joint only microns thick. A reader who understands the cold plate understands why thermal engineering is now an accelerator design problem rather than a facilities problem. > [!note] The one-sentence version > Heat leaves a processor through a stack of thin layers, and the whole art is making every layer as thin and as conductive as possible. ## Why it exists An accelerator package in current production dissipates on the order of a kilowatt through a surface the size of a playing card. Heat leaving that surface must cross into something that carries it away. Air, even moving quickly, is a poor conductor and offers very little surface area per unit volume. Liquid in direct contact with a metal block bonded to the package carries heat orders of magnitude more effectively, which is what allows a kilowatt-class device to exist at all. ## Prerequisites - **Thermal resistance:** the temperature rise produced per unit of heat flow through a path; the whole design is a campaign to lower it. - **Heat flux:** heat per unit area, measured in watts per square centimeter; a modern accelerator die operates at fluxes once associated with rocket nozzles. - **Thermal interface material:** the paste, pad, or solder filling the microscopic gaps between the chip lid and the plate. Air in those gaps is an insulator, so this layer governs the whole stack. - **Microchannel:** a fine internal passage, often fractions of a millimeter across, which multiplies the wetted surface area and thins the boundary layer where heat transfers into the fluid. - **Pressure drop:** the resistance the plate presents to flow; finer channels transfer heat better and demand more pumping energy, which is the central trade. - **Direct-to-die or bare-die cooling:** removing the package lid and bonding the plate to the silicon itself, eliminating one interface entirely. ## How it works Coolant enters the plate, spreads through a manifold into a field of microchannels directly above the hottest region of the die, absorbs heat as it passes, and leaves a few degrees warmer. The plate body is copper for conductivity, sometimes with a nickel plating for corrosion resistance, and its internal geometry is designed against a specific die's heat map rather than as a uniform grid: channels are made denser over the hot spots. Clamping is engineered as carefully as the hydraulics, because the force must be high enough to thin the interface layer and even enough to avoid stressing the package; springs and load plates distribute it to specification. Across a rack, plates are fed in parallel from manifolds so that each sees its design flow, and connected with self-sealing quick disconnects so a tray can be removed without draining anything; see [[wiki/Direct-to-Chip Liquid Cooling|Direct-to-Chip Liquid Cooling]]. Coolant chemistry is an operational discipline in its own right: corrosion inhibitors, biocide where applicable, filtration, and periodic analysis, because the loop's enemies are galvanic corrosion, particulate, and biological growth rather than anything dramatic. ## Stack position - **Upstream:** the [[wiki/Coolant Distribution Unit|coolant distribution unit]] and the facility loop behind it. - **Downstream:** the accelerator, whose sustained clock speed is set by how well this path works; a better plate is literally more performance. - **Function and vendor:** the function is unavoidable at current densities; plates are a fast-commoditizing assembly with many suppliers, increasingly shipped pre-installed as part of the tray. - **Degraded case:** the device throttles to protect itself, delivering a fraction of the performance that was purchased, and the capital in the rack underperforms quietly rather than failing visibly. ## Lineage and trajectory Water-cooled processor frames existed in mainframes of the 1960s through 1980s, and the thermal conduction module of that era is the direct ancestor of the modern plate. The practice retreated to supercomputing and enthusiast niches when chips ran cool, and returned in this decade because packages crossed the kilowatt class. The trajectory runs toward bare-die attachment, plates co-designed with the package from the start, two-phase operation where the coolant boils in the channels, and ultimately cooling structures etched into the silicon itself; several of those are in research and early deployment as of October 2026. ## Relationships - **fed by:** [[wiki/Coolant Distribution Unit|Coolant Distribution Unit]] · [[wiki/Direct-to-Chip Liquid Cooling|Direct-to-Chip Liquid Cooling]] - **cools:** [[wiki/Accelerator|Accelerator]] · [[wiki/Host CPU|Host CPU]] - **installed in:** [[wiki/Compute Tray|Compute Tray]] - **alternative to:** [[wiki/Air Cooling|Air Cooling]] · [[wiki/Immersion Cooling|Immersion Cooling]] - **source articles:** [[articles/Vertically Integrating an AI Superpower from AI Factory to Citizen|Vertically Integrating an AI Superpower from AI Factory to Citizen]] · [[articles/The Gigawatt Is the New Measure of War|The Gigawatt Is the New Measure of War]] - **collection route:** [[collections/American Infrastructure|American Infrastructure]] · [[collections/American|American Collection]] ## Evidence posture - **Established:** thermal resistance and interface materials as the governing constraint; microchannel geometry and the heat-transfer against pressure-drop trade; copper construction with corrosion-resistant plating; clamping force specification; parallel manifold feeding with self-sealing disconnects; coolant chemistry management against corrosion, particulate, and biological growth; accelerator packages in the kilowatt class. - **Established:** copper cold plates cooling accelerators in current production rack-scale systems. - **Plausible:** two-phase microchannel operation and cooling structures integrated into silicon; both are demonstrated in research and limited products, with volume deployment still to come. ## Sources - [Helios Is AMD's First AI System To Rival Nvidia Vera Rubin](https://www.cnbc.com/) — CNBC, October 2026; copper cold plates on accelerators in a current rack-scale system. - Thermal engineering literature on microchannel cooling and interface materials, as of October 2026.