Frore claims its LiquidJet can drop Nvidia Rubin GPU temperatures by 10°C — can also boost performance by 15% as hyperscalers eye using delidded GPUs in production environments

Frore Systems, a company known for its cooling solutions produced with semiconductor-grade tools, has developed a technology that can reduce the temperature of next-generation AI accelerators and boost their performance by 15%. Their LiquidJet coldplate technology claims to lower Nvidia Rubin GPU junction temperatures by up to 12°C, which could result in a 10% to 25% increase in token generation per watt.

This improvement is attributed to the fact that leakage current increases exponentially with temperature, doubling for every 10°C rise in maximum junction temperature. As leakage current grows, transistor switching becomes less efficient, causing hotter GPUs to require higher voltages to sustain clocks. This leads to Dynamic Voltage and Frequency Scaling (DVFS) reducing clocks to remain within thermal limits, ultimately resulting in decreased performance and token generation.

Nvidia designs its platforms around a maximum allowable junction temperature (Tj(max)), which is a fundamental design constraint for GPUs, packages, and cooling solutions. If the GPU can operate below this temperature, it can sustain higher clocks or lower voltage. Frore's analytical thermal model suggests that delidding the Rubin package can decrease junction temperature by up to 20°C, potentially improving tokens/Watt by up to 35%.

However, delidding GPUs may decrease mechanical reliability. Despite this, cloud system providers are exploring the use of delidded Rubin GPUs to increase token generation, despite the risks. Frore's coldplate technology utilizes manufacturing techniques from semiconductor fabrication, such as etching and bonding, to create intricate three-dimensional copper microstructures that address hotspots on the accelerator's silicon. These microstructures cannot be produced using traditional machining processes like skiving.