Frore LiquidJet Diamond Targets Nvidia Rubin’s 2,300W Cooling Problem

Frore Systems has introduced Frore LiquidJet Diamond, a custom coldplate designed to cool 2,300-watt Nvidia Rubin accelerators and unusually concentrated chip hotspots. The company claims the design can lower GPU temperatures by 16°C to 22°C compared with conventional skived coldplates. It also projects 30% to 35% more AI tokens per watt.

Those numbers would make cooling a direct contributor to computing output, rather than a supporting system that merely prevents overheating. However, Frore’s published results remain company claims based partly on thermal models and specified operating conditions. Independent Rubin hardware tests have not yet established the same gains.

The more important contest is between two cooling architectures. Conventional coldplates spread coolant across machined channels, while Frore concentrates flow around a processor’s mapped hotspots. LiquidJet Diamond adds diamond heat spreaders where heat flux is highest, placing the expensive material only where Frore believes it matters.

That approach arrives as Nvidia and AMD prepare accelerators with higher package power, larger multi-die assemblies, and denser memory systems. Cooling vendors now need to manage both total heat and small regions that become much hotter than the surrounding package.

Frore LiquidJet Diamond Puts Diamond Above the Hotspots

Frore’s central idea is to engineer the coldplate around each processor’s heat map, then place diamond directly over its most difficult hotspots.

A coldplate is a metal cooling component mounted against a processor package. Liquid circulating inside the plate carries heat toward a facility cooling loop. The thermal path usually includes the silicon, packaging materials, a thermal interface material, and the coldplate itself.

Every layer adds thermal resistance. That resistance raises the temperature difference between the coolant and the working silicon. The effect becomes harder to control when one small region produces much more heat than the rest of the package.

Frore says LiquidJet Diamond uses a three-dimensional, ultra-short-loop, multistage structure. Its internal channels direct coolant toward hotspots first, before routing it across the GPU die, high-bandwidth memory, and other chiplets.

The company manufactures the internal copper structure using processes adapted from semiconductor fabrication. This method supports shapes that are more complex than the straight fins commonly produced through skiving, a process that cuts thin channels into copper.

Diamond forms the second part of the design. Frore embeds diamond material into the coldplate’s bottom surface and positions it above identified hotspots. Diamond conducts heat more effectively than copper, allowing it to spread concentrated heat before that heat enters the liquid-cooled structure.

This is not a generic plate intended for every accelerator. Dimensions and internal channels change with the target processor. Frore lists a reference size of 104.6 by 62 by 20.4 millimeters for an Nvidia GB300 configuration.

The custom design creates the product’s most important distinction from Frore’s LiquidJet Nexus. Nexus combines several coldplates and cooling components into a larger server-level assembly. LiquidJet Diamond is an individual accelerator coldplate that system builders can integrate into their own platforms.

Frore’s published product specifications say the plate supports a 770 W/cm² hotspot measuring five by five millimeters. That rating assumes 25°C inlet coolant, a 95°C maximum junction temperature, PTM7950 thermal interface material, and an unlidded package.

Those conditions matter. The 770 W/cm² figure describes heat concentrated in a small hotspot, not the average density across an entire GPU package. It also does not mean that any accelerator can sustain that density under different coolant temperatures or packaging choices.

The company compares LiquidJet Diamond with multi-inlet skived coldplates. It claims more than 150% greater maximum hotspot capacity, 75% higher heat-transfer efficiency, and GPU temperatures between 16°C and 22°C lower.

Frore also says those temperature reductions translate into 30% to 35% more AI tokens per watt. That is a projected computing-efficiency improvement, not a measured reduction in total data-center energy use.

The launch reporting notes that Frore has not disclosed how it manufactures or integrates the diamond pieces. That missing detail could affect cost, manufacturing yield, reliability, and production volume.

LiquidJet Diamond therefore represents a specific engineering proposal. It combines localized heat spreading with coolant paths designed for one processor’s thermal behavior. Its value depends on whether those advantages survive production tolerances, long operating periods, and complete rack deployments.

Nvidia Rubin Turns Cooling Into a Compute Constraint

A 2,300-watt accelerator makes temperature control part of the performance budget, because thermal resistance can limit sustained clocks and useful output.

Traditional server cooling was often designed around average package power. That approach becomes less reliable when a large accelerator combines multiple compute dies, high-bandwidth memory, fast interconnects, and uneven workloads.

An AI workload can keep tensor-processing hardware active for long periods. Memory and interconnect activity can also shift heat across the package. A coldplate must handle both persistent total power and local spikes without allowing one region to reach its limit.

Frore plans to show LiquidJet Diamond with a 2,300-watt Nvidia Rubin cooling configuration. The demonstration targets hotspots reaching the company’s stated 770 W/cm² level.

It remains unclear whether the demonstration uses a shipping Rubin accelerator, preproduction hardware, or a thermal test vehicle. A test vehicle can reproduce power and hotspot placement, but it cannot represent every mechanical and workload behavior of production silicon.

Nvidia’s published HGX specifications place eight Rubin GPUs on an HGX Rubin NVL8 baseboard. Each GPU includes 288 GB of HBM4 memory and 22 TB/s of memory bandwidth. The eight-GPU platform provides two terabytes of HBM4.

Those specifications help explain why cooling has become a system design problem. Heat must be removed from a dense arrangement containing GPUs, memory, networking components, power delivery hardware, and processors.

Nvidia says Rubin provides 50 petaflops of NVFP4 inference performance per GPU. Its Rubin architecture includes 336 billion transistors, 224 streaming multiprocessors, and 896 Tensor Cores.

The architecture also uses a chiplet design and high-bandwidth package connections. That combination creates several potential heat sources within a limited physical area. Average package temperature cannot fully describe those local conditions.

Temperature affects output because modern processors manage frequency and voltage within electrical and thermal limits. A cooler chip can retain higher operating frequencies when another limit does not intervene.

However, lower temperature does not automatically guarantee a fixed throughput improvement. Performance also depends on workload behavior, memory access, networking, software efficiency, utilization, and power management.

Frore links cooling directly to tokens per watt because inference operators sell or consume useful output. If two systems use similar electrical power, the system producing more tokens has better compute efficiency.

That commercial framing raises the standard for evidence. Buyers need workload-level measurements across representative models, batch sizes, sequence lengths, and utilization patterns. A thermal model alone cannot show the entire operational result.

Cooling power must also be included. Pumps, coolant distribution units, facility heat exchangers, and optional chillers consume energy. A chip-level efficiency gain can shrink when the full cooling loop requires substantial additional power.

Frore argues that lower coldplate resistance can reduce dependence on colder inlet water. Warmer coolant can simplify heat rejection and improve facility efficiency. That claim is plausible, but the net result depends on climate, system design, and facility operating temperatures.

The pressure therefore falls on conventional coldplate suppliers and server designers. They must either improve hotspot control or prove that less specialized plates deliver sufficient performance at lower cost and complexity.

Conventional Coldplates Face a Hotspot Problem

The main contest is not diamond against copper alone; it is hotspot-specific cooling against coldplates designed around broader heat removal.

A skived coldplate starts with a copper base and a field of thin fins. Coolant moves between those fins, collecting heat before leaving the plate. Manufacturers understand the process, and the resulting components fit established server supply chains.

This architecture can work well when heat is reasonably distributed. It becomes less efficient when coolant travels through long channels while one small location produces extreme heat.

Designers can add inlets, change fin geometry, raise flow rates, or reduce coolant temperatures. Each response has consequences. Higher flow can increase pressure requirements, while colder liquid can require more facility energy and condensation controls.

Frore’s short-loop design attempts to reduce the distance between an inlet and the hottest region. Its multistage routing gives the coolest available liquid to the hotspot before serving other package components.

The diamond layer addresses a related problem. Coolant channels remove heat after it reaches the coldplate. A high-conductivity spreader can distribute concentrated heat across more cooling area before that transfer occurs.

This distinction explains why Frore cites both heat-transfer efficiency and hotspot density. One metric describes how effectively heat crosses the cooling structure. The other describes how concentrated the source can become before junction temperature reaches the specified ceiling.

The design also highlights a limit of comparing products through total wattage. Two accelerators consuming identical power can require different coldplates when their die layouts, packaging, and hotspot locations differ.

That is why Frore names Nvidia Rubin, Rubin Ultra, Feynman, AMD Instinct MI455, MI500, and custom hyperscaler accelerators as targets. Each would receive a plate tailored to its dimensions and heat map.

AMD’s public MI400 roadmap scheduled the CDNA-based MI400 family for 2026. Newer AMD planning also places MI400 inside rack-scale systems with HBM4 and next-generation networking.

Supporting multiple vendors could broaden Frore’s opportunity. It also multiplies the engineering work. Every new package can require updated simulations, new internal structures, tooling changes, and fresh mechanical qualification.

A standard skived plate offers a different advantage. Its familiar construction can reduce qualification risk and support multiple product configurations with fewer custom manufacturing steps.

LiquidJet Diamond must therefore earn its complexity. A buyer will compare the added thermal headroom against component cost, development time, serviceability, reliability, and supply availability.

Diamond itself does not settle that comparison. The amount, quality, orientation, bonding method, and placement of the material all affect performance. An excellent bulk conductivity number does not guarantee a low-resistance assembled interface.

Contact pressure presents another challenge. The coldplate must maintain reliable contact across a large package without damaging dies, bridges, interposers, or surrounding components.

That problem becomes more serious for unlidded packages. Removing an integrated heat spreader shortens the thermal path, but it exposes delicate package structures to mechanical loading.

Frore’s 770 W/cm² specification assumes an unlidded package. Buyers evaluating a lidded production GPU cannot transfer that figure directly to their own systems.

The conventional route therefore remains credible when it meets the accelerator’s operating requirements. Frore’s design becomes most attractive where existing plates force lower clocks, colder coolant, greater pumping power, or insufficient hotspot margin.

The market will not choose between the two approaches through conductivity figures alone. It will choose through qualified systems that deliver predictable performance across thousands of servers.

Why Lower Temperatures Can Increase Tokens per Watt

Frore’s efficiency claim rests on a clear mechanism, but the size of the gain depends on more than the coldplate.

A processor’s junction temperature reflects coolant temperature, generated heat, and total thermal resistance. Reducing resistance lowers junction temperature when the other variables remain stable.

Lower temperatures can reduce leakage power in semiconductor devices. They can also give a processor more room inside its thermal limit, supporting higher sustained frequency or reducing throttling.

Frore’s earlier LiquidJet modeling attributed a temperature reduction of roughly 6°C to 12°C to the base coldplate. The company associated that range with a 10% to 25% increase in tokens per watt.

LiquidJet Diamond adds a localized spreader that Frore says provides up to another 10°C of hotspot reduction. Its complete comparison then reaches a stated 16°C to 22°C temperature improvement and 30% to 35% more tokens per watt.

Those ranges should not be combined as universal values. Different package designs, interfaces, inlet temperatures, flow rates, and workloads can produce different outcomes.

Independent research supports the broader relationship between liquid cooling and sustained computing performance. A 2026 liquid cooling study compared liquid-cooled and air-cooled eight-GPU Nvidia H100 systems.

The researchers reported GPU temperatures between 41°C and 50°C for the liquid-cooled system. The air-cooled configuration ranged from 54°C to 72°C under load.

They also measured 54 TFLOPS per GPU on the liquid-cooled platform, compared with 46 TFLOPS per GPU using air cooling. That represented a 17% performance difference in their tested systems.

The study does not validate Frore LiquidJet Diamond or its Rubin claims. It uses different hardware and compares liquid cooling with air cooling, rather than two liquid coldplate designs.

It does show why a cooling vendor can reasonably discuss computational output. Thermal stability can affect sustained processor behavior, not merely component survival.

Tokens per watt remains a workload-dependent measurement. A memory-bound inference task might gain less from additional GPU frequency than a compute-bound task.

Software can also dominate the result. Quantization, batching, speculative decoding, kernel selection, and parallelism influence how many tokens a system produces from its electrical input.

Rack networking adds another variable. Rubin systems use high-bandwidth NVLink connections so GPUs can operate as a coordinated compute fabric. Communication overhead can limit workload performance even when individual GPUs remain cool.

The most convincing test would measure the same Rubin system with two coldplates. Both configurations would use identical software, coolant conditions, workloads, and facility accounting.

Results should include junction temperatures, hotspot variation, accelerator power, pump power, coolant flow, pressure drop, throughput, and tokens per joule. Long-duration runs would reveal whether early gains remain stable.

Data-center buyers would also need performance distributions, not only peak results. A plate that reduces variation between accelerators could improve cluster scheduling even if its average advantage appears modest.

That consistency matters because large training and inference jobs can wait for their slowest participants. A handful of thermally constrained GPUs can reduce the efficiency of a larger synchronized system.

Frore’s 35% figure is therefore best read as the upper end of the company’s projected outcome. It is not yet a universal improvement available to every Rubin deployment.

The Claims Still Need Production-Scale Proof

LiquidJet Diamond addresses a credible engineering problem, but its published numbers do not yet provide a complete production qualification case.

The first uncertainty concerns validation. Frore presents detailed conditions for its hotspot rating, yet it has not published a broad independent benchmark using production Rubin systems.

Analytical thermal models are valuable during development. They let engineers compare materials, flow paths, package choices, and coolant conditions before final hardware becomes widely available.

Models still depend on their inputs. Assumptions about power distribution, interface quality, heat-spreader contact, workload behavior, and processor controls can change the result.

The planned demonstration can provide useful evidence if Frore identifies the test device and measurement method. A thermal test vehicle would validate heat removal under controlled conditions. It would not validate token throughput.

A production GPU test could connect temperature changes with real performance. Even then, buyers would need workload details and complete system-power accounting.

The second uncertainty is manufacturing. Frore says it adapts semiconductor processes to fabricate copper structures, but it has not disclosed production cost or output capacity.

Embedding diamond introduces more interfaces and process steps. Thermal cycling can stress bonds because different materials expand at different rates.

Servers repeatedly heat and cool during maintenance, workload changes, and failures. Qualification must show that the coldplate retains contact, flow performance, and structural integrity through those cycles.

The third issue is customization. Tailoring coolant channels to each heat map can improve performance, but it ties a specific plate to a package configuration.

Late accelerator changes could require a plate redesign. That risk matters when GPU vendors revise power maps, mechanical tolerances, package heights, or interface requirements before volume shipment.

Cloud operators also value replaceability. A highly customized component needs dependable spare inventory and clear service procedures across the deployment’s operating life.

The fourth question concerns coolant infrastructure. A coldplate does not operate independently. It connects to manifolds, pumps, coolant distribution units, heat exchangers, sensors, and facility loops.

Pressure drop affects pump requirements. Flow distribution becomes more difficult when many accelerators share a rack. Small manufacturing differences can also influence which branch receives the most coolant.

Frore describes LiquidJet Diamond as a drop-in upgrade, but that phrase requires careful interpretation. Mechanical compatibility does not guarantee identical hydraulic behavior or unchanged facility requirements.

The fifth uncertainty is economics. Lower temperatures can create value through higher throughput, lower cooling energy, or greater hardware longevity. Buyers must compare those savings with the plate’s complete installed cost.

Diamond material does not automatically make a system uneconomic because Frore uses it only around hotspots. However, the company has not published enough commercial information to calculate a deployment payback.

The sixth issue is packaging risk. Frore’s highest hotspot specification assumes direct contact with an unlidded package using PTM7950.

Removing the lid reduces thermal resistance by eliminating layers between the silicon and coldplate. It can also increase the danger of cracking or uneven loading across multi-die assemblies.

An integrated heat spreader protects the package and distributes mechanical force. Hyperscalers might accept greater packaging risk for higher output, but enterprise buyers may prioritize conservative serviceability.

These questions do not invalidate the architecture. They define the work between a convincing demonstration and a qualified product deployed across an AI cluster.

Frore can strengthen its case by publishing independent test results, detailed operating conditions, reliability data, and hydraulic measurements. Customer deployments would provide stronger evidence than another modeled comparison.

Three Signals Will Determine Whether Diamond Cooling Scales

The next stage is about verification, accelerator qualification, and adoption, not another increase in the headline temperature claim.

The first signal is a transparent Rubin demonstration. Frore should identify whether the 2,300-watt setup uses production silicon, preproduction hardware, or a heater-based test vehicle.

A useful demonstration would disclose coolant temperature, flow rate, pressure drop, package configuration, interface material, and hotspot measurement method. Real workload results should include full system power.

Independent measurements matching Frore’s 16°C to 22°C range would strengthen the case for hotspot-targeted cooling. A result limited to modeled or simulated hardware would leave the main performance claim unresolved.

The second signal is formal qualification with an accelerator or server vendor. A named design win would indicate that LiquidJet Diamond passed mechanical, hydraulic, thermal, and reliability reviews.

Qualification for Nvidia Rubin or an AMD Instinct platform would also test Frore’s customization model. It would show whether the company can turn processor heat maps into repeatable production components.

A generic compatibility statement carries less weight. Buyers need evidence that the plate works inside a complete server with approved mounting forces, connectors, manifolds, firmware controls, and maintenance procedures.

The third signal is cluster-scale operating data. One coldplate can perform well while a populated rack exposes flow-balancing, pump, vibration, leak-detection, or service problems.

Long-duration results should track accelerator temperatures, performance variation, cooling power, failure rates, and workload output. Tokens per watt should include the supporting cooling equipment.

If Frore sustains its claimed efficiency improvement under those conditions, coldplate selection becomes part of AI capacity planning. Operators could extract more work from a fixed electrical allocation.

If gains shrink after system power and production packaging are included, LiquidJet Diamond may remain valuable for extreme hotspots without replacing conventional plates broadly.

The timing still favors serious experimentation. Nvidia’s Rubin platform combines high compute density with HBM4 and rack-scale connectivity. AMD is pursuing similarly integrated accelerator systems.

Meanwhile, AI operators face electrical and facility limits that cannot be solved through chip design alone. Every layer between the transistor and cooling plant now affects available computing capacity.

Frore LiquidJet Diamond offers a technically coherent response. It routes coolant according to the processor’s heat map and uses diamond where copper alone faces its hardest local load.

The remaining question is whether that precision survives manufacturing, qualification, and rack-scale operation. Buyers should watch measured Rubin workloads, named system partners, and independently verified energy accounting before treating 35% more tokens per watt as an established result.