Exclusive: Nvidia Approves Samsung’s HBM3 Chips for China-Market CPUs

Samsung's HBM3 Chips Receive Conditional Greenlight from Nvidia

Samsung Electronics' latest high bandwidth memory (HBM3) chips have been conditionally approved by Nvidia for use in its processors, marking the first such approval, according to three insider sources. However, this approval is limited; the HBM3 chips will initially be integrated only into the less advanced Nvidia H20 GPU, developed specifically for the Chinese market in line with U.S. export restrictions. It remains uncertain whether Samsung's HBM3 chips will be employed in Nvidia's other AI processors, or if further testing will be required for broader use.

Current Status and Future Prospects of HBM3 and HBM3E Chips

Samsung is yet to meet the requirements for Nvidia’s next-generation HBM3E chips, with testing still ongoing. Both companies have refrained from commenting on the matter. High bandwidth memory, a form of dynamic random access memory (DRAM) that stacks chips vertically to save space and reduce power usage, is critical for GPUs involved in artificial intelligence due to its ability to handle large datasets from complex applications.

Nvidia's approval comes amid a surge in demand for advanced GPUs, driven by a boom in generative AI. This high demand has strained supply, with only three main manufacturers - SK Hynix, Micron, and Samsung - producing HBM. Nvidia is eager for Samsung to fulfill its standards to diversify its supply base. The need for diverse suppliers is pressing as SK Hynix, the leader in HBM, plans to shift more of its production to HBM3E, thereby reducing HBM3 output.

Challenges and Progress in Meeting Nvidia’s Standards

Samsung, the world's leading memory chip maker, has been striving to meet Nvidia's standards for both HBM3 and HBM3E chips since last year. Challenges related to heat and power consumption have been significant, although Samsung has publicly denied issues stemming from these factors.

The introduction of Samsung’s HBM3 for Nvidia's H20 processor is anticipated as early as August. The H20, designed for the Chinese market, is part of Nvidia's compliance with tightened U.S. export restrictions intended to limit advancements in supercomputing and AI that could have military applications. This GPU, however, has its computing power capped compared to the more advanced H100, which is available outside China.

The Market Dynamics

Initially, the H20 saw subdued demand when it launched, priced below a competing chip from China's Huawei. Nonetheless, sales have started to pick up swiftly, as confirmed by separate sources. SK Hynix remains Nvidia's primary HBM supplier, having provided HBM3 chips since June 2022 and recently starting HBM3E shipments in March. While Micron has also committed to supplying Nvidia with HBM3E, exact deployment strategies and timelines remain undisclosed.

Final Thoughts

Samsung’s incremental progress in securing Nvidia approval for its HBM3 chips marks a vital step as the company seeks to expand its footprint amidst escalating demand for advanced memory solutions crucial to AI and supercomputing. Whether Samsung can overcome its challenges and meet Nvidia’s rigorous standards for broader application in future Nvidia processors will be crucial in the high-stakes battle of memory chip supremacy.