Source: Online Compilation
△ Scan to read
Following NVIDIA’s announcement at the end of August regarding the mass production of the Rubin architecture chip next year, NVIDIA discussed the progress of Rubin at the Goldman Sachs Technology Conference on September 8. NVIDIA CFO Colette Kress stated that the Rubin chip is already preparing for market entry, and the Rubin architecture will consist of six chips, all of which have entered tape-out. This news has garnered widespread attention in the semiconductor and artificial intelligence fields, signaling another significant leap for NVIDIA in chip technology.
The Rubin chip is manufactured using TSMC’s advanced 3nm EUV process and is equipped with HBM4 high-bandwidth memory, initially configured with 8 stacks. This process and memory configuration result in impressive performance parameters for the chip. Its floating-point computing capability reaches 1 petaflop per second, representing a significant improvement of about 30% compared to previous products. Each GPU will have 288GB of video memory, with bandwidth soaring from the previous 8TB/s to 13TB/s. Such powerful memory and bandwidth capabilities can better meet the demands of complex artificial intelligence computations and large-scale data processing.
In the data center sector, the emergence of the Rubin chip is expected to reshape the landscape. NVIDIA plans to launch the next-generation flagship AI server, the NVIDIA Vera Rubin NVL144 CPX, based on the Rubin architecture. This server will integrate 36 Vera CPUs, 144 Rubin GPUs, and 144 Rubin CPX GPUs. Test data shows that the Rubin rack equipped with Rubin CPX can outperform the current flagship rack, the GB300 NVL72, by up to 6.5 times when processing large context windows. The next-generation flagship rack will provide 8 exaFLOPs of NVFP4 computing power, 7.5 times higher than the GB300 NVL72, while a single rack can offer 100TB of high-speed memory and 1.7PB/s of memory bandwidth. This means that data centers using Rubin chip-related architectures can process massive amounts of data more efficiently, making it highly attractive for cloud service providers and enterprises engaged in large-scale data analysis.
From the perspective of artificial intelligence applications, the powerful performance of the Rubin chip also brings new possibilities for various application scenarios. For example, in the field of natural language processing, current models often face insufficient computing power when processing longer texts, while the Rubin chip, with its strong floating-point computing capability and high-bandwidth memory, is expected to enable language models to handle longer and more complex texts, achieving more accurate semantic understanding and text generation. In computer vision, the Rubin chip can significantly reduce processing time and improve processing accuracy for high-definition video processing and large-scale image dataset analysis.
Additionally, NVIDIA plans to launch the Rubin Ultra in the second half of 2027, further enhancing GPU performance and introducing HBM4E memory technology. By then, the FP4 inference computing capability will reach up to 15 exaFLOPS, and FP8 training performance will also reach 5 exaFLOPS. The new NVL576 layout can accommodate up to 576 GPUs, supporting enterprises in building ultra-large-scale computing clusters, which will provide a solid hardware foundation for cutting-edge research in artificial intelligence, such as the exploration of general artificial intelligence.
The successful tape-out of NVIDIA’s next-generation Rubin chip marks an important milestone in its chip development journey. With the mass production of the Rubin chip next year and the subsequent launch of products like Rubin Ultra, it is expected to drive multiple fields, including artificial intelligence and data centers, into a new stage of development globally, further consolidating NVIDIA’s leading position in the semiconductor and artificial intelligence computing fields.
Welcome to Submit Articles
The “Semiconductor Chip Technology” media platform is open for submissions from readers!
There are no restrictions on topics or genres. If your company has the latest product releases,
cutting-edge technological advancements, or favorable news regarding financing and mass production,
you are welcome to submit via the following email~
Recommended by Industry Professional Media
Clickto read the original article
