2026-01-28 · 供应链分析

2027-28年GPU散热瓶颈或致AlSiC材料替代铜铝。

涉及标的： $NVDA

中文翻译

散热可能是2027-2028年GPU的一个问题：随着散热目标向2000-3000W+范围攀升，铜和铝变得不足。但AlSiC（铝硅碳化物），一种金属基复合材料，可能变得重要。原因如下： Rubin VR200 GPU指向1800W TDP，AMD Instinct MI450X据报道高达2500W，$NVDA Rubin Ultra的功耗据报道达到2300W。向2300W的转变创造了大规模的热通量问题。 AlSiC（铝硅碳化物）是一种金属基复合材料，用于国防、太空和高功率工业应用（如高铁）。就像Toto等随机马桶公司成为HBM关键一样，这种用于航空航天和工业散热的材料复合材料可能会被用于AI，因为它可以在不发生分层的情况下承受数万次热循环。因此，随着Rubin等AI加速器达到与工业电源模块或太空相当的功率水平，半导体中AlSiC的采用可能变得重要。对于2300W系统中的AlSiC： - （结到壳）：包括硅芯片、底部填充物和第一层TIM（热界面材料） - （壳到散热器）：包括均热板（顶盖）和第二层TIM（TIM2） - （散热器到环境）：冷板或液体换热器的热阻。壳到散热器可能是AlSiC的应用场景。随着功率密度可能继续向3000W攀升，可能会出现机械和散热危机，这可能导致材料从传统铜封装发生变化。所以这里有四个部分： 1. 内部/外部散热看起来SiC（碳化硅）中介层（内部）可能用于$NVDA Rubin代GPU。然后AlSiC（外部）可能用于微通道顶盖（MLCP）或位于芯片中介层组件上方的均热板。 2. 3D垂直堆叠（SoIC）这些复杂的封装在热循环期间极易翘曲。SiC中介层是脆性晶体，无法提供结构刚性。AlSiC充当“加强筋”，防止基板在高夹紧压力下弯曲。 3. Rubin Ultra NVL576机架可能达到高千瓦功率密度这种密度造成了SiC中介层无法解决的重量负载瓶颈。Rubin Ultra NVL576机架散热堆栈加液体流道的累积重量可能超过地板承重限制（AlSiC可能成为必要，以在不牺牲传热性能的情况下将散热管理的“死重”减少60%以上）。 4. 微通道的净成型制造传统铜顶盖必须蚀刻或CNC加工，据报道这一过程在Rubin量产中遇到了高难度。 AlSiC使用“快速注射成型”制造陶瓷预成型体，然后渗入铝。这允许创建复杂的内部几何形状。AlSiC微通道顶盖和硅集成IHS看起来是铜散热管理的替代品。我们可能会在2026年初看到这一问题的解决，因为SemiAnalysis在2025年报道Nvidia的Blackwell（B100/B200）因CoWoS-L封装中的翘曲问题面临良率问题。 TLDR散热可能是2027-2028年的瓶颈：一些潜在受益者可能是SiC中介层（高纯度SiC粉末）和用于2027-2028年散热的AlSiC复合材料。这都是持续的研究，但也许我们会在1-2年内看到一些极其小众且随机的铁路或太空AlSiC供应商被AI使用，就像HBM的马桶制造商一样。

英文原文

Thermal is a likely issue for 2027-2028 GPUs: As thermal targets escalate toward the 2000-3000W+ range, copper and aluminum becomes insufficient. But, AlSiC, a metal matrix composite, may become important. Here's why: Rubin VR200 GPU pointed toward 1800W TDP, AMD Instinct MI450X, reportedly reaches up to 2500W, and $NVDA Rubin Ultra's power reportedtly goes to 2300W. The move toward 2300W creates a heat flux problem of massive scale. Aluminum Silicon Carbide (AlSiC) is a metal matrix composite, used as for defense, space, and high-power industrial applications (eg. high-speed rail). Similar to how random toilet companies like Toto became critical for HBM, this material composite used for thermal management for aerospace and industrial might be used for AI as it can survive tens of thousands of thermal cycles without delamination. So, as AI accelerators like Rubin reach power levels comparable to industrial power modules or space, the adoption of AlSiC in semiconductors may become important. For AlSiC in 2300W systems there's: - (Junction-to-Case): Includes the silicon die, the underfill, and the first layer of TIM - (Case-to-Sink): Includes the heat spreader (lid) and the second layer of TIM (TIM2) - (Sink-to-Ambient): The thermal resistance of the cold plate or liquid heat exchanger. Case-to-Sink is likely the application for AlSiC. And as power densities will likely continue to climb toward the 3000W mark there becomes a mechanical and thermal crisis that likely causes a material change from traditional copper packaging. So there's four different parts to this: 1. Internal/External Thermal It does look like SiC interposers (internal) are probably used for $NVDA Rubin gen GPUs. Then AlSiC (external) may be used for the Microchannel Lid (MLCP) or heat spreader that sits on top of the die interposer assembly. 2. 3D vertical stacking (SoIC) These complex packages are highly vulnerable to warping during thermal cycling. A SiC interposer is a brittle crystal and it cannot provide structural rigidity. AlSiC acts as a "stiffener" that prevents the substrate from bowing under the high clamping pressures 3. Rubin Ultra NVL576 rack likely reaches high KW of power density This density creates a weight-loading bottleneck that SiC interposers cannot solve. Rubin Ultra NVL576 rack cumulative weight of the thermal stack plus liquid manifolds can exceed the floor limits (and AlSiC may become necessary to reduce the "dead weight" of the thermal management by 60%+ without compromising heat transfer). 4. Net-Shape Manufacturing of Microchannels Traditional copper lids must be etched or CNC-machined, a process that has reportedly encountered high mass-production difficulty for Rubin volumes. AlSiC is manufactured using "Quickset Injection Molding" to create a ceramic preform that is then infiltrated with aluminum. This allows for the creation of complex internal geometries. AlSiC Microchannel Lids and Silicon Integrated IHS looks like the alternatives for copper for thermal management. We might be seeing this addressed earlier in 2026 as SemiAnalysis in 2025 reported that Nvidia’s Blackwell (B100/B200) faced yield issues specifically due to warpage in the CoWoS-L packaging. TLDR thermal is a likely a bottleneck in 2027-2028: Some beneficiaries are potentially SiC interposers (high-purity SiC powder) and AlSiC composite for thermal management in 2027-2028. This is all ongoing research, but maybe we'll see some extremely niche and random small railway or space AlSiC supplier be used up for AI in 1-2 years time like toilet makers for HBM.

在 X 上查看原推 ↗