Frore LiquidJet liquid cooling cuts Nvidia Rubin GPU temps by 10°C, boosts token-per-watt by 15%

Frore Systems, a maker of semiconductor-grade cooling solutions, published a white paper demonstrating that its LiquidJet coldplate technology can reduce NVIDIA Rubin GPU junction temperatures by up to 10–12°C, translating to a 10–25% improvement in tokens-per-watt and a 15% boost in token generation in analytical models. The company projects that holistic cooling-stack improvements—spanning GPU packaging, thermal interface materials (TIMs), coldplates, and coolant temperatures—can increase token generation per watt by more than 30%.

The efficiency gains are rooted in physics: leakage current approximately doubles for every 10°C increase in junction temperature, and transistor switching power rises by ~2% over the same range. Modern AI accelerators like Rubin dissipate up to 2,400 W and can hit 95°C+ die temperatures, triggering dynamic voltage and frequency scaling (DVFS) that reduces clocks, performance, and token output. Lowering operating junction temperature improves efficiency even when the processor does not hit its thermal ceiling, because reduced leakage and switching losses directly translate to higher sustained throughput.

Frore's analysis highlights thermal resistance as a critical bottleneck: delidding (removing the integrated heat spreader) on a Rubin GPU can reduce junction temperature by up to 20°C compared to stock, potentially improving tokens/watt by 35%. Some hyperscalers are already exploring delidded Rubin GPUs in production environments to maximize token economics, though the approach trades mechanical reliability for performance gains.

For practitioners: in AI datacenters where cost-per-token drives unit economics, a 15% improvement in token generation directly impacts TCO and ROI. Cooling is no longer just a thermal management problem—it is a workload profitability lever. NVIDIA-centric infrastructure teams should evaluate LiquidJet and similar coldplate upgrades as part of capacity planning for next-gen accelerators.