Frore Systems says its LiquidJet coldplate could cut Nvidia Rubin GPU junction temperatures by 6°C to 12°C, a reduction the company models as enough to improve tokens per watt by 10% to 25%.

Tom's Hardware reported the claim from a Frore white paper that treats cooling as a direct performance and efficiency variable for AI accelerators, not just a way to keep chips below safe operating limits.

The figures are not measured results from deployed Rubin systems.

They come from an analytical thermal model built around the equation Tj = Tinlet + Q × Rtotal, using published or assumed Rubin operating parameters and Frore's estimates for coldplate thermal resistance.

Frore uses a 95°C junction-temperature assumption for Rubin in the analysis.

It says leakage power roughly doubles for every 10°C increase in junction temperature, while transistor switching power rises by about 2% over the same range.

That matters because hotter GPUs may need higher voltage to sustain clocks, and dynamic voltage and frequency scaling can then reduce frequency to stay inside thermal and power limits.

The model breaks total thermal resistance into the GPU package, the thermal interface material and the coldplate.

Frore's central claim is that lowering resistance at those layers can raise token-generation efficiency even when a GPU is not already throttling.

LiquidJet Claim

LiquidJet is Frore's coldplate design made with semiconductor-style etching and bonding rather than conventional skiving.

The company says the approach allows short microchannels around hotspots, multiple cooling stages and flow routing matched to the GPU's power-density map.

For Nvidia Rubin, Frore models that design as reducing junction temperature by 6°C to 12°C.

A roughly 10°C reduction would correspond to about a 15% increase in token generation efficiency, according to the white paper.

Frore also argues the design changes facility-cooling economics.

Nvidia designed Rubin to operate with coolant entering at up to 45°C, allowing many AI data centres to rely on free cooling without mechanical chillers.

In Frore's example, a Rubin GPU using a conventional skived coldplate needs a chiller coefficient of performance of about 6.7 before colder coolant provides a net efficiency benefit.

LiquidJet lowers that break-even point to about 4.1.

Tom's Hardware also reported Frore's view that LiquidJet is more efficient on Rubin data-centre GPUs than on Blackwell data-centre GPUs because Rubin has higher transistor density.

Delidding Trade-Off

The white paper goes further than Frore's own coldplate.

It says delidding Rubin, by removing the integrated heat spreader and the graphene thermal interface material between the die and lid, could reduce junction temperature by as much as 20°C compared with a lidded package.

Frore models that as improving tokens per watt by up to 35%.

The mechanical risk is the constraint.

Without the heat spreader, a Rubin package built with TSMC's CoWoS-L technology would be more vulnerable to cracking in the bridges connecting the two Rubin dies.

Uniform contact pressure across multiple exposed dies is also harder to maintain than on a monolithic processor.

Tom's Hardware noted that even in the Hopper era, some GPUs cracked with certain liquid coolers.

Frore says some cloud system providers are exploring delidded Rubin GPUs to improve token generation and output, but the report does not name those providers or cite production deployments.

Nvidia's Rubin reportedly uses liquid indium metal thermal interface material with gold-plated contact surfaces to reduce the thermal penalty of a lidded package.

Frore argues that an unlidded package paired with a high-performance phase-change material such as PTM7950 would still have lower overall thermal resistance than a lidded package using liquid metal, giving as much as a 14°C junction-temperature advantage and up to a 28% gain in tokens per watt in the company's model.

The strongest published result remains modelled rather than field-tested: Frore claims LiquidJet can cut Rubin junction temperatures by 6°C to 12°C and improve tokens per watt by 10% to 25%, while delidding and thermal-interface changes could push the modeled efficiency gain as high as 35% at greater mechanical risk.