Frore LiquidJet Diamond: Diamond cold plate to cool AI GPUs by up to 22°C
As modern AI accelerators reach kilowatt-level power consumption, managing intense hotspots has become a critical bottleneck for cooling. To address this, Frore Systems introduced its LiquidJet Diamond technology. By combining microchannel fluid routing with integrated diamond heat spreaders, the company claims it can reduce GPU temperatures by 16–22 °C compared to conventional cold plates.
Architecture: Microchannels and Diamond Spreaders
The LiquidJet Diamond builds upon the LiquidJet architecture introduced in 2025. This design employs a multi-stage microchannel structure with optimized flow paths, ensuring that coolant reaches high-density heat zones before moving to the rest of the chip. This targeted approach is designed to mitigate thermal loads precisely where they are most intense.
The standout feature is the addition of diamond structures integrated into the underside of the cold plate, positioned directly over the chip's hottest regions. Since diamond possesses significantly higher thermal conductivity than copper, it can dissipate concentrated thermal loads much more efficiently. This is vital because local hotspots often limit a GPU's maximum power draw and sustained boost clocks, even if the average chip temperature remains manageable.
Frore Systems specifies that the technology can handle local heat flux densities of up to 770 W/cm². These specifications are based on a 5 × 5 mm hotspot, a coolant inlet temperature of 25 °C, a maximum chip temperature of 95 °C, and the use of Honeywell PTM7950 thermal interface material.
Performance and Compatibility
According to manufacturer data, the diamond layer provides an additional 10 °C reduction over the original LiquidJet solution, leading to a total temperature improvement of 16–22 °C. Frore Systems also claims a 30–35% increase in computational efficiency (Tokens per Watt). However, it is important to note that lower temperatures do not automatically guarantee higher throughput without considering the specific power limits and thermal management of the accelerator itself.
The technology is aimed at high-performance computing environments. At the OCP Global Summit, Frore demonstrated the tech in conjunction with an NVIDIA Rubin accelerator (rated at 2,300 W). The company also stated that the technology is intended for future accelerators like AMD Instinct MI455, MI500, Rubin Ultra, and custom AI ASICs.
Currently, these temperature and efficiency gains are based on manufacturer comparisons and have not yet been verified by independent testing. For data center operators, the practical impact will depend on whether the improved heat transfer is offset by increased requirements for coolant flow, pressure, and overall system energy consumption.
Гика
