Key points
- Samsung reportedly turns to TSMC for Nvidia's memory
- What Nvidia told its three memory suppliers
- Samsung's backup plan if TSMC runs short
Samsung Electronics (005930) plans to use TSMC's (TSM) 3-nanometer process to make the base die for Nvidia's (NVDA) custom high-bandwidth memory, called NVHBM, the Seoul Economic Daily reported Sunday. Samsung and TSMC have previously worked together on HBM packaging compatibility. This would be their first collaboration on manufacturing the chips themselves, according to the report.
Nvidia asked Samsung, SK Hynix (SKHY), and Micron (MU) to use TSMC's 3nm process for their NVHBM base dies and the same DRAM generation for the memory chips stacked above them. Samsung, Nvidia, and TSMC did not comment in the reports.
What changes with NVHBM
HBM combines several DRAM chips, known as core dies, in a stack above a logic chip called the base die. The base die connects the memory stack to the processor.
With NVHBM, Nvidia moves its custom memory controller from the processor into that base die. The company introduced the design in August.
Nvidia says that gives up to 30% more memory bandwidth than standard HBM4E and cuts HBM power use by 15%. It also frees up to 25% more area on the compute die of an XPU, the custom AI processor the memory connects to. Amazon's (AMZN) Annapurna Labs is the first to work on NVHBM, starting with its Trainium4 chip.
Why Nvidia wants one process
When each supplier makes the memory a different way, it's hard to get the same performance from every stack, and Nvidia has to check each supplier's version separately, industry sources told the Seoul paper. Nvidia asked for core dies on 1c, the sixth generation of 10-nanometer-class DRAM.
Samsung's chip division accepted the terms after an internal review.
Micron already uses a foundry for the next generation. "On HBM4E, this co-design is on a foundry process, both for the customized product and for the JEDEC-specified product," Scott DeBoer, Micron's technology chief, said on the company's Sept. 30 earnings call. Micron used its own base die for HBM4.
What Samsung keeps in-house
Samsung's HBM4, which it began mass-producing earlier this year, and its HBM4E, now in mass-production testing, use a base die from Samsung's own 4nm foundry process. Samsung will keep making both the memory and the base die itself for standard HBM and for other custom customers, including AMD (AMD).
Samsung is also testing whether its own foundry can make NVHBM base dies in case Nvidia's demand outgrows TSMC's capacity. A follow-up report said that backup would use Samsung's 2nm process. The paper cited industry estimates putting TSMC's 3nm yields above 90% and Samsung's 2nm yields around 60%. Those are estimates for two different processes, and they may involve different chip designs, so they aren't a direct comparison.
Samsung may not be able to cover custom HBM demand alone. Its 4nm base-die production is limited to Line 17 in Hwaseong and part of its P3 plant in Pyeongtaek, the follow-up said.
UBS forecasts 2027 HBM demand of about 61.5 billion gigabits, up 91% from this year, against supply of 54.6 billion gigabits, according to the follow-up report. That would leave a shortfall of about 7 billion gigabits.
On the same call, Manish Bhatia, Micron's chief operating officer, said the company expects custom HBM "to be a high-value product as well." He added, "we're confident that HBM will continue to contribute and be a strong ROI product for us."
