As Samsung Electronics prepares to supply high bandwidth memory (HBM) at scale for AMD's next-generation AI platform "Helios," the two companies said they achieved industry-leading sixth-generation HBM (HBM4) performance through AI memory co-design collaboration.
Samsung Electronics on Aug. 23 (local time) held a separate session at the AMD Advancing AI 2026 event at Moscone West in San Francisco and disclosed the results of its cooperation with AMD. Bart Walker, vice president of the Samsung Electronics Device Solutions (DS) Division's Memory Business, and principal engineer James Ahn took the stage as presenters.
◇ 20-year partnership… "Becoming the primary supplier in the AI era"
Vice President Walker traced the starting point of the partnership to 2007. He said the partnership began when Samsung Electronics supplied GDDR4 for AMD's Radeon HD 2000 series. Since then, cooperation has expanded with 14-nanometer FinFET process collaboration in 2014, a Radeon intellectual property (IP) partnership spanning server SSDs and DRAM in 2019, and SmartSSD co-development in 2020.
In 2021, they jointly developed the industry's first CXL DRAM, and in 2023 expanded the IP license to develop Exynos 2400 based on AMD's RDNA3, broadening collaboration across graphics, mobile, and computing. In 2025, 12-high stacked HBM3E was installed in AMD's Instinct MI350 series AI accelerators.
Walker emphasized in materials presented on-site that Samsung Electronics solidified its status as AMD's primary supplier by signing a new contract in Mar. this year. He cited AMD CEO Lisa Su's visit to Samsung Electronics' Pyeongtaek campus early this year as the backdrop for the expanded cooperation. The two companies signed a strategic cooperation agreement for next-generation AI memory solutions at the time, and Su was quoted as saying, "To realize next-generation AI infrastructure, deep collaboration across the industry is necessary."
◇ HBM4 in Helios secures performance and efficiency with a 4-nanometer base die
James Ahn, principal engineer at Samsung Electronics, said, "Samsung Electronics' HBM4 will be installed in AMD's latest Instinct MI455X AI accelerator and the Helios platform," adding, "The Instinct GPU lineup will feature HBM4 and HBM3E; the server CPU EPYC lineup will use the RDIMM DRAM module; and the Helios platform will be supplied with both RDIMM and HBM4."
HBM4 doubles the number of input/output (I/O) pins from 1,000 pins (1K) in HBM3 and HBM3E to 2,000 pins (2K). Ahn said, "Despite the increased I/O count, we applied a FinFET structure based on Samsung Electronics Foundry's 4-nanometer process to the base die to improve power efficiency." He said this delivers more than 200% higher bandwidth, up to 70% smaller area, and up to 50% lower power consumption compared with the conventional planar structure.
HBM4 currently in mass production and supply supports a 36-gigabyte (GB) capacity with 12-high stacking. The next-generation HBM4E will be based on a 32-gigabit (Gb) 1c core die and is set to support up to 64GB per stack, up to 16 gigabits per second (Gbps) per pin, and up to 4 terabytes (TB) of bandwidth per stack, as well as up to 16-high stacking. This improves performance by 20%, power efficiency by 16%, and scaling flexibility by 14% over HBM4.
Samsung Electronics is also supplying DRAM modules (RDIMM) for next-generation EPYC CPUs supported on the Helios platform. Ahn said, "We developed a 32Gb DDR5 using the industry's first 12-nanometer-class process," adding, "This product delivers up to 8.0 Gbps speeds, 1.25 times higher performance than 6.4 Gbps RDIMM. Capacity scales to 256GB using 32Gb TSV (through-silicon via), which is twice that of 32Gb monodie. Power consumption for 32Gb 2Rx4 at 6.4 Gbps is reduced by 15% compared with 16Gb 2H."
Ahn added, "The architecture underpinning AI ultimately determines performance, and next-generation DRAM and NAND play a key role in both total cost of ownership (TCO) and performance," noting, "Samsung Electronics and AMD are working closely to advance technologies that enable AI and data center workloads."