As Nvidia unveiled the next-generation AI chip Rubin CPX, specialized for advanced artificial intelligence (AI) functions such as video production and software development, analysts say supply of the latest graphics DRAM GDDR7 (7th-generation GDDR) from the memory chip industry will gain momentum. That is because GDDR7 will be installed not only in Nvidia's latest gaming graphics card, the GeForce RTX 5090, but also in Rubin CPX, a lower-spec AI accelerator designed for the Chinese market.
On the 9th (local time), Nvidia unveiled Rubin CPX, based on the design of the next-generation AI chip Rubin Platform to be released next year. Nvidia said it plans to ship Rubin CPX by the end of this year at the latest. Nvidia said, "AI models need up to 1 million tokens to process one hour of video content, which was difficult to achieve with existing graphics processing units (GPUs)," and noted, "(Through Rubin CPX) it can be used for software coding and video production that use millions of tokens." A token is the smallest unit of data created by breaking down information so AI can learn, generate, and infer.
The Rubin CPX unveiled this time is Nvidia's first AI chip that can implement not only coding for software development, but also high-quality video production and the ability to understand long user prompts and provide the necessary answers. Jensen Huang, Nvidia's chief executive officer (CEO), said, "Rubin CPX is the first GPU specially designed for large-context AI (Context AI), where models simultaneously infer millions of knowledge tokens." Context AI is an AI technology that identifies a user's situation, conversation, and external information—the context—in real time to provide more accurate and personalized answers.
As Nvidia releases chips equipped with GDDR7 one after another, analysts expect supply from the memory chip industry to expand. In January this year, Nvidia released the GeForce RTX 5090 gaming graphics card equipped with GDDR7. GDDR7 is applied to the device, and Samsung Electronics is known to have virtually monopolized the initial supply. It is also reported to have signed the supply contract for products in that series ahead of competitors, not just the RTX 5090. SK hynix and Micron are also said to have begun shipping GDDR7 to Nvidia.
Nvidia is also reported to be planning to adopt GDDR7 for the low-cost AI chip B40 targeting the Chinese market. Nvidia is designing a new AI chip with reduced performance due to U.S. export controls to China. Among the options under consideration is lowering performance by applying GDDR7, which has lower bandwidth—a measure of data transfer speed—than high bandwidth memory (HBM). Considering demand in the Chinese market, the industry expects at least 1 million units of Nvidia's B40 chips to ship this year and up to 5 million next year.
The Rubin CPX disclosed by Nvidia will be equipped with 128 gigabytes (GB) of GDDR7, not HBM. With the Rubin CPX launch set for late next year, memory chip corporations are expected to begin mass production in the first quarter at the latest and start supplying Nvidia. A semiconductor industry official said, "It is true that demand for GDDR7 is increasing sharply," and added, "As Samsung Electronics and SK hynix are preparing for mass production in line with Nvidia's product release roadmap, supply is expected to increase."