Blab Tech
Samsung Unveils HBM5 and zHBM Roadmap, Targeting 200-% and 800-% Performance Gains
At the heart of the latest semiconductor sprint, Samsung Electronics took center stage at SEMICON Taiwan on September 1, 2026, unveiling a bold high‑bandwidth memory (HBM) roadmap that promises to reshape the AI and high‑performance computing (HPC) landscape.
The company introduced two next‑generation products: HBM5 and a later‑stage zHBM. Samsung’s presentation projects HBM5 to deliver up to a 200 % performance increase over the current HBM4E standard, while zHBM is expected to deliver roughly eight times the performance of HBM4E. The figures refer to memory bandwidth and efficiency—metrics that are critical for AI accelerators and HPC systems. Samsung added that HBM5 will improve performance per watt by 20 % and reduce thermal resistance by 20 % compared with HBM4E. The zHBM design, which stacks memory directly above AI accelerators, is projected to triple performance per watt and cut thermal resistance by up to 90 %. These gains are aimed at easing the memory bottleneck that industry analysts say is tightening as AI workloads outpace HBM bandwidth growth.
To guide its development, Samsung introduced the “CUBE” strategy—Capacity, Utilization, Bandwidth, and Efficiency. According to the company, CUBE drives its work on 3‑dimensional memory structures, targeting higher capacity and bandwidth while improving power efficiency and thermal performance. Samsung clarified that the approach extends beyond HBM to other memory products, including DRAM, NAND flash, and enterprise solid‑state drives.
The roadmap was illustrated with a mock‑up of an HBM5 module that appeared at Computex 2026. The mock‑up features a heat‑path block cooling solution and a 2 nm base die, underscoring Samsung’s commitment to advanced process nodes for memory. In a separate Korean media report, Samsung said it has secured 70 % of its memory production capacity for long‑term agreements, with those deals priced at roughly five times lower than spot prices—an effort to lock in customers amid the AI‑driven surge in memory demand.
Samsung is one of the three dominant HBM manufacturers, alongside SK Hynix and Micron. The JEDEC standard for HBM4 was approved in April 2025, and HBM5 is slated to follow in the coming years. Samsung first disclosed high‑level details of its HBM roadmap at the Hot Chips 2026 conference. At SEMICON, executive vice president for memory product planning Choi Jang‑seok explained the CUBE strategy in depth.
For AI chip designers, the promised bandwidth and power savings could translate into smaller cooling systems and denser integration of AI accelerators. The zHBM architecture, which places memory directly on top of a GPU or AI core, could also simplify board design and reduce latency.
Samsung’s focus on 3‑D stacking aligns with a broader industry trend toward vertical integration of memory and logic. The 2 nm base die for HBM5 indicates that Samsung is preparing to leverage its advanced semiconductor manufacturing capabilities to meet next‑generation AI demands.
At the time of the announcement, Samsung had not yet provided a commercial launch date for HBM5 or zHBM. The company is expected to release detailed specifications and production timelines in the coming months, likely in coordination with AI chip partners. Analysts predict that the new memory technologies will become commercially available in the late 2020s.
In summary, Samsung’s September 2026 roadmap positions the company to lead the next wave of high‑bandwidth memory. The announced performance gains—up to 200 % for HBM5 and eightfold for zHBM—are designed to meet the escalating bandwidth and power efficiency requirements of AI and HPC systems. The CUBE strategy and advanced 3‑D structures signal Samsung’s continued investment in memory innovation, while the secured long‑term contracts indicate strong market demand.
The company introduced two next‑generation products: HBM5 and a later‑stage zHBM. Samsung’s presentation projects HBM5 to deliver up to a 200 % performance increase over the current HBM4E standard, while zHBM is expected to deliver roughly eight times the performance of HBM4E. The figures refer to memory bandwidth and efficiency—metrics that are critical for AI accelerators and HPC systems. Samsung added that HBM5 will improve performance per watt by 20 % and reduce thermal resistance by 20 % compared with HBM4E. The zHBM design, which stacks memory directly above AI accelerators, is projected to triple performance per watt and cut thermal resistance by up to 90 %. These gains are aimed at easing the memory bottleneck that industry analysts say is tightening as AI workloads outpace HBM bandwidth growth.
To guide its development, Samsung introduced the “CUBE” strategy—Capacity, Utilization, Bandwidth, and Efficiency. According to the company, CUBE drives its work on 3‑dimensional memory structures, targeting higher capacity and bandwidth while improving power efficiency and thermal performance. Samsung clarified that the approach extends beyond HBM to other memory products, including DRAM, NAND flash, and enterprise solid‑state drives.
The roadmap was illustrated with a mock‑up of an HBM5 module that appeared at Computex 2026. The mock‑up features a heat‑path block cooling solution and a 2 nm base die, underscoring Samsung’s commitment to advanced process nodes for memory. In a separate Korean media report, Samsung said it has secured 70 % of its memory production capacity for long‑term agreements, with those deals priced at roughly five times lower than spot prices—an effort to lock in customers amid the AI‑driven surge in memory demand.
Samsung is one of the three dominant HBM manufacturers, alongside SK Hynix and Micron. The JEDEC standard for HBM4 was approved in April 2025, and HBM5 is slated to follow in the coming years. Samsung first disclosed high‑level details of its HBM roadmap at the Hot Chips 2026 conference. At SEMICON, executive vice president for memory product planning Choi Jang‑seok explained the CUBE strategy in depth.
For AI chip designers, the promised bandwidth and power savings could translate into smaller cooling systems and denser integration of AI accelerators. The zHBM architecture, which places memory directly on top of a GPU or AI core, could also simplify board design and reduce latency.
Samsung’s focus on 3‑D stacking aligns with a broader industry trend toward vertical integration of memory and logic. The 2 nm base die for HBM5 indicates that Samsung is preparing to leverage its advanced semiconductor manufacturing capabilities to meet next‑generation AI demands.
At the time of the announcement, Samsung had not yet provided a commercial launch date for HBM5 or zHBM. The company is expected to release detailed specifications and production timelines in the coming months, likely in coordination with AI chip partners. Analysts predict that the new memory technologies will become commercially available in the late 2020s.
In summary, Samsung’s September 2026 roadmap positions the company to lead the next wave of high‑bandwidth memory. The announced performance gains—up to 200 % for HBM5 and eightfold for zHBM—are designed to meet the escalating bandwidth and power efficiency requirements of AI and HPC systems. The CUBE strategy and advanced 3‑D structures signal Samsung’s continued investment in memory innovation, while the secured long‑term contracts indicate strong market demand.
