NVIDIA (NVIDIA) is emerging in the artificial intelligence (AI) chip market. At the GTC annual technology conference held in March this year, Huang Renyi revealed the chip blueprint for three years, allowing customers to be more at ease to maintain...
NVIDIA (NVIDIA) is emerging in the artificial intelligence (AI) chip market. At the GTC annual technology conference held in March this year, Huang Renyi revealed the chip blueprint for three years, allowing customers to be more at ease to maintain long-term cooperation with NVIDIA, and also highlights that NVIDIA's demand for advanced power packaging of the crystalline foundry manufacturer will become increasingly greater.
Huang Renfeng predicts the blue picture of the third generation chip, including the Blackwell Ultra NVL72, which is launched in the second half of 2025 and uses the Blackwell GPU (picture processor) architecture; the Vera Rubin NVL144, which is expected to be launched in the second half of 2026 and uses the Rubin architecture; the Rubin Ultra NVL576, which is expected to be launched in the second half of 2027 and uses the Rubin architecture; and the Feynman architecture chip that has not yet been released in the details of the 2028.
Huang Renqiu said that technology companies often regard development route diagrams as highly confidential, but he believes that NVIDIA is not only a technology company, but also an AI-based facility company. It must describe the development route diagram in detail so that people around the world can refer to it.
Looking at NVIDIA's positive AI chip GB300, through advanced packaging technology, two Blackwell GPUs and high-frequency wide memory are packaged together to achieve stronger Blackwell Ultra chips, and then use ultra-high-speed networks to connect 72 Blackwell Ultra, which improves the overall efficiency by 50%.
NVIDIA plans to launch Vera Rubin architecture chips in 2026, which are considered as Blackwell's advanced version, introducing the new HBM4 high-frequency wide memory and the 6th-generation NVLink interconnection technology, which can provide faster data transmission and GPU connection.
Vera Rubin architecture cabinet can connect to 144 GPUs, and the number of high-level series links reaches 576 GPUs. These technologies demonstrate NVIDIA's strategy to accelerate product iteration cycles, and also highlight the increasing demand for advanced NTIE packaging.
With the computing power of new generation GPUs such as Blackwell and Rubin, the high-speed data transfer between thousands of GPUs has become a huge challenge. Traditionally, through the copper cable's signal transmission, encountered problems such as power dissipation and limited frequency width density, began to excited the CPO (co-packaged optical component) technology based on silicon photons, and built the originally pluggable external optical fiber receiver module directly next to the exchanger chip.
NVIDIA has been demonstrated at the GTC Conference that the next generation of Spectrum-X ethernet platform will introduce CPO technology. By tightly integrating optical components with interchanger chips, it can provide ultra-high frequency width and reduce power consumption on a single chip, which helps the AI data center improve data transmission stability, reduce operation costs and overcome hot-emission challenges.
NVIDIA's investment in CPO silicon photonic technology means that it is no longer just a company that sells GPU chips alone, but provides a complete solution from computing, internal interconnection to external data transfer. It not only consolidates NVIDIAAI's dominance, but also affects the future trend of high-efficiency computing and data center industries.
Extended reading: Silicon photonic key technology: optocoupling, Taiwan factory layout
