2026.09.23 | AC | Thermal | Thermal Material Expert: Rubin Graphene

摘要

Rubin TIM2 and TIM1 material solutions, rationale behind the final selection of graphene, and subsequent upgrade roadmaps. Rubin graphene material value and suppliers, and domestic leaders' technical capabilities. Optical module thermal materials.

The following content reflects expert sharing for reference only and does not constitute any investment advice.

Q&A

What are the specific differences in the TIM2 thermal solutions between GPUs and CPUs on the NVIDIA Rubin platform, and what is the value of the graphene pads adopted?

The NVIDIA Rubin platform has finalized the adoption of a graphene pad solution. For GPUs, which have a copper lid, the graphene pad serves as TIM2, filling the interface between the copper lid and the cold plate. The CPU does not feature a copper lid, so the graphene pad is applied directly between the die and the cold plate, functioning as a TIM1.5 application.

While the graphene pads share the same material specifications across both scenarios, differences in dimensions and thickness lead to different values. The pad used for the GPU measures approximately 65mm x 55mm; priced at 30,000 to 50,000 yuan per square meter, each pad is worth about 130 to 140 yuan. The CPU pad is slightly smaller at roughly 55mm x 45mm, but its thickness is increased to align flush with the cold plate, resulting in a unit value of about 70 to 80 yuan. For rough estimation purposes, a standardized value of 100 yuan per pad can be applied.

Why did the NVIDIA Rubin platform ultimately choose graphene pads over liquid metal as its thermal material, and how sustainable is this solution across future product iterations?

Initially, the industry expected the Rubin platform to stick with the indium-gallium-tin alloy liquid metal solution that NVIDIA deployed on its 5,090 reference graphics card. Although liquid metal outperforms graphene in initial thermal conductivity, practical testing revealed several issues. The first is reliability: after prolonged operation at high temperatures, liquid metal experiences significant performance degradation due to oxidation, alongside the risk of corroding contact metals such as copper and aluminum. In contrast, while graphene pads have slightly lower initial thermal performance, they demonstrate high stability and reliability.

The second issue is assembly efficiency. Graphene pads can be pre-cut to shape and rapidly mounted via robotic arms with vacuum suction cups, taking only a few seconds per unit. In contrast, applying liquid metal involves a complex process requiring precise control over thickness and uniformity, additional anti-corrosion requirements such as gold or nickel plating on the cold plate, and the installation of sealing rings to prevent leakage and isolate air. Consequently, its overall assembly efficiency is far lower than that of graphene pads.

Based on reliability and assembly efficiency considerations, the Rubin platform ultimately adopted graphene pads. For the next-generation Rubin Ultra, this solution will most likely be retained because these core issues are difficult to resolve in the short term. Therefore, this thermal design is expected to persist across at least several product generations, including Rubin and Rubin Ultra.

As future chip power consumption climbs from 2,300W to 5,000W or even higher, is there a performance ceiling for graphene pads?

The suitability of graphene pads does not depend strictly on absolute chip power consumption. Their current application spectrum is broad, spanning from optical modules consuming just 20 to 30 watts to Rubin chips reaching up to 2,000 watts. When selecting thermal solutions—especially for server chips requiring 24/7 continuous operation—reliability is often prioritized over absolute peak thermal conductivity.

In terms of thermal resistance, liquid metal (indium-gallium-tin alloy) is an order of magnitude lower than graphene pads, traditional thermal grease, and phase change materials. Graphene pads have thermal resistance on par with grease and phase change materials, but offer superior reliability under high power without the performance degradation seen in thermal grease from silicone oil volatilization or in phase change materials from pump-out effects. Therefore, even if future chip power consumption reaches 4,000W or 5,000W (such as on the Feynman platform), as long as liquid metal's reliability issues remain unresolved, customers may still favor graphene pads. They can compensate by optimizing other parts of the cooling system (such as upgrading cold plates or adopting diamond-copper lids) to ensure long-term system stability.

What are the respective unit values of the four TIM materials—thermal grease, phase change materials, graphene pads, and liquid metal—in single-chip applications?

Taking the chip dimensions of the Blackwell Rubin platform as a benchmark, the unit values across different TIM materials diverge significantly. If high-end thermal grease is used (such as Dow products with thermal conductivity of 6W/m·K, priced at roughly 2,000 to 3,000 yuan/kg), the dispensing volume per chip is about 1 to 2 grams, costing just one or two yuan. If high-end phase change sheets from the previous-generation GB200/GB300 are used (priced at roughly 2,000 to 3,000 yuan/m²), the unit cost is about 8 to 9 yuan.

The value of a graphene pad is significantly higher—roughly 10 times that of a phase change sheet and 50 times that of thermal grease—reaching the 100-yuan level per unit. Liquid metal currently costs around 20 to 30 yuan per chip, which is lower than graphene. However, prices for the primary metal raw materials in liquid metal, such as indium, gallium, and tin, continue to climb, whereas graphene pad costs are trending downward as suppliers multiply and technology matures. Liquid metal costs could potentially surpass those of graphene pads within the next two to three years.

What are the application forms of diamond materials in the TIM (thermal interface material) field, and will they replace existing TIM materials such as graphene?

Diamond materials have four primary application forms in thermal dissipation. In the short to medium term, none directly substitute mainstream TIM materials like graphene; rather, they serve as complements or are deployed in other sections of the cooling system.

First, as thermally conductive fillers. Adding diamond powder to conventional thermal grease or gel improves thermal conductivity, effectively broadening the application scenarios for traditional TIMs.

Second, as diamond-copper composites. These replace pure copper chip lids to enhance heat conduction from the die to the lid, though TIM2 is still required on top of the lid.

Third, as chip substrates. Used in optical modules or GaN power semiconductors as the underlying substrate, with TIM materials still required on top for thermal transfer.

Fourth, as diamond heat spreaders. These can be attached directly onto bare dies as TIM1, but still require TIM1.5 or TIM2 above them to interface with the heat sink.

Therefore, none of these diamond applications directly replaces the TIM layer itself, posing no threat to the market for graphene or other conventional TIM materials.

What is the technical feasibility and cost of bonding diamond heat spreaders as TIM1 to the chip, and does the chip lid need to be retained?

Under current assessments, using diamond heat spreaders as TIM1 still requires retaining the chip lid, unless a single diamond heat spreader can cover all dies—which is economically unfeasible. Diamond heat spreaders carry prohibitive costs; covering a die area as large as the Rubin platform could cost several thousand yuan per unit. Under current technological and cost structures, customers with a thermal budget of 1,000 to 2,000 yuan have multiple mature, cost-effective alternatives. This approach would only be considered regardless of cost if all other cooling solutions fail. Consequently, commercial adoption of this approach is unlikely within the next three years due to excessive costs.

What are the commercial prospects for diamond heat spreaders?

Diamond heat spreaders face limited commercial room in the near term. The primary reason is that their cost far exceeds that of liquid metal, diamond-copper, or existing microchannel technologies, without necessarily offering superior thermal performance. Balancing cost against performance, commercial adoption appears very difficult to achieve in the short term.

Who are the main competitors in supplying graphene pads for the NVIDIA Rubin platform?

There are two main competitors: domestic Changzhou Fuxi and Germany's Henkel overseas. The expectation that market share can remain high through 2027 is based on the following: Changzhou Fuxi's core business is supplying graphene raw materials, and aggressive expansion into finished pads would cannibalize its raw material sales, dampening its expansion appetite. Henkel, an incumbent supplier on earlier NVIDIA generations, has not conducted deep internal R&D on this type of TIM; instead, it partnered with Sweden's Smart High Tech and, backed by NVIDIA, only began investing in facility construction in June 2026. Given that moving from sampling to volume production and yield ramp-up takes at least six months to a year, neither competitor is expected to achieve large-scale shipments in 2027, posing no meaningful disruption to the prevailing landscape.

What are the primary challenges of applying graphene materials to TIM1?

There are two main challenges. First, the application scenarios and customer touchpoints are fundamentally different. TIM1.5 and TIM2 interface with chipmakers or downstream module makers and ODMs/foundries, whereas TIM1 serves as an in-package material requiring direct engagement with OSATs like TSMC, ASE, and Amkor. Packaging architectures are driven primarily by TSMC and ASE rather than fabless designers like NVIDIA. Second, the qualification and verification cycles are exceptionally long. While TIM2 requires about a year for qualification, TIM1 demands at least two to three years. OSATs evaluate criteria far beyond basic thermal performance and reliability, such as compatibility with other packaging materials and inspectability under acoustic microscopy (e.g., C-SAM, P-SAM), where graphene pads encounter hurdles. Qualification requirements vary widely across OSATs and chip architectures, making the path to secured orders a protracted process.

What is the current adoption status of thermal materials in optical modules? Are graphene thermal pads suitable?

From 400G to 1.6T optical modules, 99.99% use thermal gel as their TIM. This is because optical modules house compact circuit boards populated with numerous uneven heat sources; only thermal gel can conformally fill these microscopic gaps while supporting automated dispensing assembly. As optical transceiver speeds advance from 800G to 1.6T and eventually

1. 2T, thermal conductivity requirements for gels have escalated from 8W/m·K to 15–16W/m·K, driving a three- to four-fold

increase in unit value. Graphene thermal pads are unsuitable here. Deploying them in optical modules presents several roadblocks: DSP chips are tiny (just 6mm x 7mm), yielding low unit value per cut pad (around 5 to 6 yuan); edge sealing is required to prevent flaking, which introduces severe reliability risks inside optical modules with stringent insulation demands due to graphene's electrical conductivity. Overall, the unfavorable risk-reward profile makes this an impractical application. Graphene pads are better suited for future NPO (Near-Package Optics) and CPO (Co-Packaged Optics) architectures, where optical engines are packaged in close proximity to switch ASICs, rather than current pluggable form factors.

Why have domestic manufacturers achieved breakthroughs in graphene thermal pads while remaining behind in other TIM materials?

Domestic players lag overseas competitors in conventional TIM materials largely due to deficits in foundational science and theoretical R&D, relying instead on supply chain integration—specifically combining upstream powders with downstream formulations. Graphene thermal pads represent the only TIM category where domestic players have broken through in recent years. This success stems not from insurmountable technical barriers, but from specific structural factors: first, raw material cost advantages, as China became the world's largest graphene production base around 2021 with the lowest global feedstock pricing; second, labor cost advantages, given that graphene pad manufacturing remains labor-intensive; and third, accumulated process know-how across multi-step flows involving extensive non-automated tooling, where competitive advantage hinges on process optimization rather than fundamental technological breakthroughs. These combined factors underpin domestic leadership in graphene pads—advantages that cannot be replicated across other material systems.

How does the value of thermal gel vary across optical modules of different data rates?

For optical modules operating at 800G and below, thermal gel accounts for roughly 3 to 4 yuan per module. For 1.6T modules, which require high-performance gels sourced from overseas vendors, the value triples or quadruples to 12 to 13 yuan per module, or even higher.