Recent Market Rumors: NVIDIA B300 HGX Series May Shift to Lidless Packaging
近期市場傳言 NVIDIA B300 HGX 系列很有可能開始「不採用」均熱片 (蓋子)，而是回歸前一代Hopper 系列用的邊緣 stiffener (加固物) Lidless 的設計，且認為B300為兩個獨立封裝的 B300A GPU 放在一塊，而非B200的Dual die結構，因此NVIDIA B300，以及CPU、GPU的Vera、Rubin皆不會採用均熱片，而是改回Lidless設計，這有助於降低 NVIDIA 生產的成本，但可能不利於 Jentech。
OSAT廠主要將晶圓代工廠生產完的晶圓切割後與載板封裝在一起，當中也會有molding、布線等半導體製程，以確保晶片能安全運作，不受外界環境汙染或影響。根據晶片需求，部分晶片最上層會加裝均熱片後完成封裝。封裝技術可分為 Ring Type（Lidless）與 Lid Type 兩大類：
Ring Type（Lidless）Package：周圍設有環狀框架 (Ring Frame) 來支撐與封裝晶片，並直接採用高效能TIM (Thermal Interface Materia) 與外部散熱模組連接。這種設計減少額外的熱組，提升散熱效果，但由於缺少均熱片及蓋子，晶片結構可靠性較低，不耐碰撞。目前NVIDIA H100、AMD MI300……等晶片皆採用此類封裝
Lid Type Package：除了環狀框架外，晶片上方還加裝均熱片，以協助熱能均勻分散，提升晶片可靠性並提供保護。
晶片散熱的關鍵在於避免單一區塊過熱，導致晶片故障。因此，需要均熱片等TIM協助有效分散熱能，類似於國中實驗課程加熱使用的陶瓷纖維網。隨著Blackwell、Rubin晶片的TDP進一步提升至超過1,000w，單點熱能密度提高，極有可能需要均熱片來協助均勻分散熱點。
此外，NVIDIA H100 與 B100 的封裝形式有所不同：
H100 採用 TSMC CoWoS-S，連接 1 顆 Hopper GPU 與 6 顆 HBM，並使用整片 Interposer。由於晶粒承載於完整 Interposer 上，結構相對穩定，因此可採用 Ring Type 封裝。
B100 採用 TSMC CoWoS-L，連接 2 顆 Blackwell GPU 與 8 顆 HBM，並使用片斷化的 Interposer，導致結構穩定性降低。
根據我的研究，CoWoS-L封裝下，Lid Type相較Ring Type的平整度多出20%，若採用Ring Type，後續ODM廠在SMT 過程中的良率將大幅下降，縱使後續有機會展出具成本優勢的Ring Type封裝晶片，但要實現大規模量產仍具挑戰。此外，後續NVIDIA Rubin、Vera晶片預計為2顆N3 晶粒與1顆N4晶粒組成，可能仍需均熱片來提升結構強度，但在正式 Tap Out 之前仍存變數。
另一個散熱關鍵是導熱效率。晶片的熱必須有效的傳導至散模組散熱，才能發揮最佳散熱效能。目前常見的TIM材質為散熱膏、散熱液、液態金屬及均熱片。均熱片主要材質為銅，近年來亦研發出石磨稀、銦的材質：
單層石磨稀散熱效果較其它材質優異許多，但需要具方向性，多層石磨稀必須排列一致，才能發揮最佳散熱效能，否則效果甚至較傳統材質差。目前僅日本少數廠商具備生產能力，但良率很低
銦材質的均熱片生產相對成熟，許多廠商具備量產能力。
另外有廠商開發半導體製程的散熱片，直接與晶片連接，以取代均熱片，但仍處於開發階段，目前銦片仍為晶片廠商的主要選擇。
我認為，目前Blackwell將採用銦片做為均熱片，主要供應商仍為Jentech，I-Chiun有機會切入成為次要供應商，並搭配All-Ring或是Horng Terng的點膠植散熱設備。
最後分享一下NVIDIA晶片的進度。根據我的研究，目前B200在台積電投片生產數量持續下降，B300投片量逐漸提升，推測2025Q1會完成產品世代交替，2025Q2 B200投片量將會趨近於0。Blackwell B200/B300 仍採用Dual die設計，由2顆GPU與8顆8/12Hi HBM 3E搭配CoWoS-L封裝，然而皆有推出Single die的降規版本B200A/B300A，B200A/B300A採用1顆GPU與4顆12-hi HBM3E die搭配CoWoS-S封裝，後續B300A改回成CoWoS-L封裝，但目前降規版本不受到市場青睞。
歡迎交流！
Recent market rumors suggest that the NVIDIA B300 HGX series is likely to stop using a heat sink (lid) and instead revert to the lidless stiffener design used in the previous Hopper series. Additionally, it is believed that the B300 consists of two separate B300A GPU packages, rather than the dual-die structure of the B200. Therefore, NVIDIA B300, as well as the Vera and Rubin CPU/GPU, will not adopt a heat sink and will instead return to a lidless design. This shift helps reduce NVIDIA's production costs, though it may negatively impact Jentech.
OSAT Packaging Process
OSAT manufacturers primarily handle the cutting of wafers produced by foundries and integrating them with substrates. This process also involves molding, wiring, and other semiconductor packaging steps to ensure that the chip operates safely without external contamination. Depending on the chip's thermal requirements, some packages include a heat sink (lid) as the final step.
Packaging technologies can be categorized into two types:
Ring Type (Lidless) Package:
Features a ring frame around the chip for support and packaging.
Directly integrates high-performance thermal interface materials (TIMs) with external cooling modules.
This design eliminates additional heat spreaders, improving thermal dissipation efficiency.
However, due to the absence of a heat sink (lid), the chip’s structural reliability is lower, making it more susceptible to damage from impact.
Currently used in chips such as NVIDIA H100 and AMD MI300.
Lid Type Package:
In addition to the ring frame, a heat sink (lid) is added on top of the chip.
This helps evenly distribute heat, improving chip reliability and providing protection.
Importance of Heat Dissipation in High-Power Chips
The key to chip heat dissipation is to prevent localized overheating, which could lead to chip failure. This is where TIMs like heat sink play a crucial role in evenly distributing heat, similar to how ceramic fiber mats are used in junior high school science experiments for uniform heat dispersion.
As the TDP of Blackwell and Rubin chips increases to over 1,000W, the rising heat density may make heat sink necessary to evenly distribute hotspots.
Differences in Packaging Between NVIDIA H100 and B100
H100:
Uses TSMC CoWoS-S.
Connects one Hopper GPU with six HBM chips.
Utilizes a full interposer, providing greater structural stability.
Can adopt a Ring Type (Lidless) packaging approach.
B100:
Uses TSMC CoWoS-L.
Connects two Blackwell GPUs with eight HBM chips.
Utilizes a segmented interposer, reducing structural stability.
CoWoS-L Packaging Challenges
According to my research, under CoWoS-L packaging, Lid Type packaging exhibits 20% better flatness compared to Ring Type packaging. If Ring Type is used, the yield rate in the SMT process at ODM factories will significantly drop.
Although a cost-effective Ring Type packaging solution may emerge in the future, achieving large-scale mass production remains a challenge.
Additionally, upcoming NVIDIA Rubin and Vera chips are expected to consist of two N3 dies and one N4 die, which may still require a heat sink to enhance structural integrity. However, there are still uncertainties before the final tap-out.
Key Factor in Heat Dissipation: Thermal Conductivity
For optimal cooling, heat must be efficiently transferred from the chip to the external cooling module. Common TIM materials include:
Thermal paste/grease
Heat transfer fluid
Liquid metal TIMs
Heat sink (heat spreaders)
Heat sink are primarily made of copper, but newer materials such as graphene and indium are also being explored:
Single-layer graphene provides superior thermal performance, but requires precise alignment for directional heat conduction. Multi-layer graphene must be properly stacked to achieve optimal thermal performance; otherwise, its performance may be worse than traditional materials. Currently, only a few Japanese manufacturers are capable of production, but yield rates remain low.
Indium-based heat sink are more mature in production, and many manufacturers already have mass production capabilities.
Some companies are developing semiconductor process-integrated heat spreaders that directly connect to the chip to replace traditional heat sink , but this technology is still in development.
Currently, indium-based vapor chambers remain the primary choice for chip manufacturers.
Blackwell’s Heat Spreader Supply Chain
Wilson Research indicates that Blackwell GPUs will use indium-based heat sink, with Jentech as the primary supplier. I-Chiun has a chance to enter the supply chain as a secondary supplier, corperate with All-Ring or Horng Terng equipment .
NVIDIA Chip Production Updates
Wilson Research also reports that B200 wafer production at TSMC is declining, while B300 wafer production is increasing. It is expected that B300 will completely replace B200 by 2025Q1, and B200 wafer production will approach zero by 2025Q2.
Both B200 and B300 utilize a dual-die design, featuring two GPUs and eight 8Hi/12Hi HBM3E chips, packaged with CoWoS-L. However, NVIDIA is also releasing downgraded , single-die versions—B200A and B300A:
B200A / B300A:
Uses one GPU and four 12-Hi HBM3E chips.
Initially used CoWoS-S packaging, but B300A later switched back to CoWoS-L.
Market reception for these downgraded versions has been weak.
Conclusion
Overall, thermal and structural considerations will continue to influence chip packaging choices.
Welcome to further discussions and insights!