AMD's Latest Memory Strategy Revealed, Samsung's "HBM + RDIMM" Dual-Track Solution Fuels Helios

According to Korean media reports, Samsung Electronics is ramping up volume supply of HBM for AMD's next-generation AI platform, Helios. The two firms have achieved industry-leading sixth-generation HBM (HBM4) performance through joint AI memory co-design.

Two Decades of Partnership Intensifies, Securing Next-Gen AI Infrastructure

Bart Walker, Senior Director at Samsung Electronics' DS Division, elaborated on Samsung's collaboration with AMD at the AMD Advancing AI 2026 event. Their partnership dates back to 2007, when Samsung supplied GDDR4 memory for AMD's Radeon HD 2000 series graphics cards. The scope of cooperation has expanded steadily since then, covering 14nm FinFET process technology, Radeon IP for server SSDs and DRAM, SmartSSD, the world's first CXL DRAM, and joint development of the Exynos 2400 processor built on AMD's RDNA 3 architecture. In 2025, Samsung's 12-layer stacked HBM3E was deployed on AMD's Instinct MI350 series AI accelerators.

In March this year, AMD CEO, Dr. Lisa Su, visited Samsung's Pyeongtaek campus. The two parties signed a strategic partnership agreement for next-generation AI memory solutions, further cementing Samsung's position as AMD's core memory supplier. Deep technical alignment enables both companies to keep pace with rising AI compute demands.

Helios Equipped with Samsung HBM and RDIMM, Delivering Comprehensive Performance and Efficiency Upgrades

AMD CEO Dr. Lisa Su officially unveiled AMD's first rack-scale AI system, Helios, at AMD Advancing AI 2026. Helios integrates AMD's latest-generation Instinct MI455X AI accelerators and sixth-generation "Venice" EPYC CPUs. A single Helios rack supports up to 72 MI455X units, delivering a total HBM4 capacity of 31TB and aggregate memory bandwidth reaching 1.7PB/s. Performance gains at the chip level can be scaled all the way to full rack deployments.

James Ahn, Senior Engineer at Samsung Electronics, explained that Samsung HBM4 will be deployed on AMD’s latest Instinct MI455X AI accelerators and the Helios platform. The architecture framework is structured as follows: the Instinct GPU family leverages HBM4 and HBM3E, EPYC server CPUs adopt RDIMM DRAM modules, while the Helios platform integrates both RDIMM and HBM4.

The latest HBM4 is stacked 12-high to deliver 432GB of capacity per chip and 23.3TB/s of bandwidth. HBM4 doubles I/O pin count versus HBM3 and HBM3E, and employs a base die fabricated on Samsung's 4nm FinFET process to boost power efficiency. Compared to the existing planar structure, it delivers over 200% higher bandwidth, 70% reduced footprint, and 50% lower power consumption.

Samsung supplies DRAM modules (RDIMM) for the next-generation EPYC CPUs supported by the Helios platform. Using its 12nm process, Samsung has developed the industry's first 32Gb DDR5 memory device. Compared with existing 6.4Gbps RDIMMs, the new product offers 125% performance with data speeds up to 8.0Gbps. Supported by 32Gb through-silicon via (TSV) technology, the maximum storage capacity reaches 256 GB, representing a doubling of capacity versus existing 32Gb single-die solutions.  Power consumption is reduced by 15% versus 16Gb 2H solution built on 32Gb 2Rx4 6.4Gbps silicon.

Ongoing R&D Targeting HBM4E, Long-Term Layout for AI Compute Memory Requirements

AMD and Samsung Electronics continue to jointly advance memory capacity and bandwidth, and their collaboration will deepen further for HBM4E, the seventh-generation HBM currently under development.

HBM4E adopts 32Gb DRAM dies, supporting a maximum single-stack capacity of 64GB, peak per-pin transfer speed of 16Gbps, and single-stack bandwidth up to 4TB/s, alongside compatibility with 16-layer stacking. Relative to HBM4, it achieves a 20% performance uplift, 16% improved energy efficiency and 14% higher scalability.