Direct-to-Chip Liquid Cooling for the Modern Data Center

Direct-to-chip liquid cooling pulls heat away from the CPU and GPU as it’s generated. Instead of moving cold air across a server, a cold plate sits directly on the silicon and a thermal fluid carries the heat out of the rack. It’s the most efficient way to cool modern accelerators, and at the densities AI workloads demand, it’s quickly becoming the only practical way. Accelsius engineers two phase direct-to-chip cooling, the next generation of the architecture, purpose-built for racks running 100kW and beyond.

A direct-to-chip system replaces the heat sink on a processor with a cold plate. Thermal fluid flows through the plate, absorbs heat from the chip, and carries it out of the server to a heat rejection unit somewhere outside the rack. The equipment itself stays dry with the fluid loop sealed and contained.

The benefit is straightforward. Air is a poor thermal conductor, and liquid is roughly 1,000 times better. By moving the heat exchange to the chip surface, direct-to-chip cooling eliminates the inefficiencies of pushing air across server fins and lets operators run silicon at full TDP without throttling, hot spots, or fan walls screaming at full speed. For background on the transition, see our piece on moving beyond the limitations of traditional data center cooling.

Not all direct-to-chip systems work the same way. The fluid choice changes everything about performance, safety, and operating cost.

This is the Accelsius approach. We break down the trade offs in Why Two Phase is the Winning Choice vs. Single Phase and look closer at the engineering in A Closer Look at Two Phase Liquid Cooling.

NeuCool is a two phase direct-to-chip liquid cooling system designed for the workloads driving the AI buildout. The numbers that matter:

• 4500W+ cooled per socket, with industry leading 0.020°C/W thermal resistance at 700W+ TDP
• 35% lower annual OpEx versus single phase direct-to-chip
• 8 to 17% lower TCO versus competing liquid cooling architectures, validated in independent Jacobs analysis
• Waterless, dielectric refrigerant with zero ozone depletion potential, ITE safe in the event of a leak
• Retrofit ready in existing facilities, no white space rebuild required

The math on air cooling stopped working somewhere around the H100. A single H100 produces over 700W. A B200 produces more. Eight per server, racked at density, and traditional CRAC and CRAH units physically cannot move enough air to keep up. Direct-to-chip cooling is what makes those deployments possible without ripping out the white space and rebuilding from scratch.

The other architectures fall short for different reasons. Air cooling can’t handle the heat. Immersion cooling handles the heat but rewrites every service procedure and fluid handling protocol in the facility. Direct-to-chip threads the needle. It keeps the rack layout that operators already know, keeps service workflows intact, and delivers the thermal performance AI silicon requires.

For more on this transition, see Embracing Direct to Chip Cooling in the AI Era and Direct to Chip Innovation Revolutionizing Data Center Cooling for AI and HPC.

NeuCool supports the processors driving today’s AI and HPC workloads.

Intel Sapphire Rapids, Intel Emerald Rapids, AMD EPYC Genoa, AMD EPYC Turin, NVIDIA Grace

NVIDIA H100, H200, B200, AMD MI325X, AMD MI355X

The system is deployed and supported through a global partner network including Equus Compute Solutions, Computacenter, and IM Data Centers, with North American based operations close to the customer.

Whether you’re retrofitting an existing facility, designing a greenfield AI build, or evaluating direct-to-chip cooling against single phase or immersion alternatives, our team will walk you through the architecture, the math, and the deployment path. Book a Demo