Samsung to use 3-nanometre process for Nvidia memory, Reuters reports
Key Facts
Reuters cited the Seoul Economic Daily as reporting that Samsung will use Taiwan Semiconductor Manufacturing Co.'s 3-nanometre process to produce base dies for Nvidia's custom high-bandwidth memory. The move concerns the "NV-HBM" project, which Nvidia announced on August 26, 2026 as part of an expansion of its "NVLink Fusion" ecosystem. According to the report, Samsung plans to retain its in-house production model for standard high-bandwidth memory. It will also continue manufacturing products customised for other customers in-house, limiting its use of the Taiwanese foundry to the Nvidia project. Samsung is also considering producing base dies for this project at its own foundry if demand exceeds the Taiwanese company's production capacity, according to Reuters.
In its announced "HBM4E" memory design, Samsung combines a dynamic random-access memory manufacturing process with a base die produced by its own foundry, according to a company statement. The memory chips use a sixth-generation process in the 10-nanometre class, known as "1c". The logic base die is manufactured at Samsung's foundry using a 4-nanometre process. The company offers integrated development and manufacturing services, spanning product design through mass production, as a way to shorten development cycles. Samsung said in its August 2026 announcement of its memory vision that these services also help improve performance and power efficiency.
Nvidia's custom memory architecture moves the memory controller into the high-bandwidth memory stack, according to the company's technical blog. The design includes integrating a custom physical-layer interface, instead of using space for the memory controller on the processor die. Nvidia described the change as "moving the memory controller into the three-dimensional memory stack and integrating a custom physical-layer interface". The company says the design frees up as much as 25% additional space on the compute die compared with standard "HBM4E" memory. It announced that Amazon's Annapurna Labs would be the first to collaborate on the custom memory technology alongside the "NVLink Fusion" interconnect ecosystem.
On performance, Nvidia says the custom memory delivers up to a 30% increase in memory bandwidth compared with standard "HBM4E" memory. It adds that it reduces high-bandwidth memory power consumption by 15% compared with the same standard memory. SK Hynix previously announced plans to use the Taiwanese company's advanced logic manufacturing process to produce the base die for "HBM4" memory. SK Hynix said using that process allows functions to be added to the base die within a limited area. Plans to use Taiwanese manufacturing technology thus cover both SK Hynix's base die and Samsung's Nvidia-specific project, according to the former's announcement and the report on the latter.
Samsung had shipped initial samples of "HBM4E" memory, according to its product announcement. In the trading session when the sample shipments were announced, the company's shares rose as much as 6.5% in morning trading, according to a Reuters report published on May 29, 2026. The report compared that with a 2.3% rise in the Korean market's benchmark Kospi index. That share price move came when the samples were announced, before the report on the use of the Taiwanese company's technology in Nvidia's custom memory. Samsung plans to begin mass production of "HBM4E" memory in line with customer schedules, after shipping initial samples and making improvements, according to its statement.
Also in "HBM4E" memory development, Micron announced that work was under way on its product based on "1-gamma" dynamic random-access memory technology. The company said it expects mass production to begin in calendar year 2027. Micron thus specifies a calendar year for expected production, while Samsung ties the start of its mass production to customer schedules. Samsung set the date of its next financial disclosure in its investor relations announcement for its third-quarter earnings conference. The conference will be held on October 29, 2026 at 10 a.m. Korea time.