Samsung Electronics, a world chief in superior reminiscence expertise, right this moment introduced that it has begun mass manufacturing of its industry-leading HBM4 and has shipped industrial merchandise to prospects. This achievement marks a primary within the {industry}, securing an early management place within the HBM4 market.
By proactively leveraging its most superior Sixth-generation 10 nanometer (nm)-class DRAM course of (1c), the corporate achieved steady yields and industry-leading efficiency from the outset of mass manufacturing — all completed seamlessly and with none further redesigns.
“Instead of taking the conventional path of utilizing existing proven designs, Samsung took the leap and adopted the most advanced nodes like the 1c DRAM and 4nm logic process for HBM4,” stated Sang Joon Hwang, Executive Vice President and Head of Memory Development at Samsung Electronics. “By leveraging our process competitiveness and design optimization, we are able to secure substantial performance headroom, enabling us to satisfy our customers’ escalating demands for higher performance, when they need them.”
Setting the Bar for Maximum Performance and Efficiency
Samsung’s HBM4 delivers a constant processing velocity of 11.7 gigabits-per-second (Gbps), exceeding the {industry} normal of 8Gbps by roughly 46% and setting a brand new benchmark for HBM4 efficiency. This represents a 1.22x improve over the utmost pin velocity of 9.6Gbps of its predecessor, HBM3E. HBM4’s efficiency could be additional enhanced as much as 13Gbps, as effectively, successfully mitigating knowledge bottlenecks that intensify as AI fashions proceed to scale up.
Also, complete reminiscence bandwidth per single stack is elevated by 2.7x in comparison with HBM3E, to a most of three.three terabytes-per-second (TB/s).
Through 12-layer stacking expertise, Samsung affords HBM4 in capacities starting from 24 gigabytes (GB) to 36GB. The firm may even preserve its capability choices aligned with future buyer timelines by using 16-layer stacking, which can broaden choices to as much as 48GB.
In order to deal with energy consumption and thermal challenges pushed by the doubling of information I/Os from 1,024 to 2,048 pins, Samsung has built-in superior low-power design options into the core die. HBM4 additionally achieves a 40% enchancment in energy effectivity by leveraging low-voltage by silicon by way of (TSV) expertise and energy distribution community (PDN) optimization, whereas enhancing thermal resistance by 10% and warmth dissipation by 30%, in comparison with HBM3E.
By bringing excellent efficiency, vitality effectivity and excessive reliability to tomorrow’s datacenter environments, Samsung’s HBM4 allows prospects to realize maximized GPU throughput and successfully handle their complete price of possession (TCO).
Comprehensive Yet Agile Production Capabilities
Samsung is dedicated to advancing its HBM roadmap by its complete manufacturing sources — together with one of many largest DRAM manufacturing capacities and devoted infrastructures within the {industry} — making certain a resilient provide chain to fulfill the projected surge in HBM4 demand.
A tightly built-in Design Technology Co-Optimization (DTCO) between the corporate’s Foundry and Memory Businesses permits it to safe the best requirements of high quality and yield. Additionally, intensive in-house experience in superior packaging permits for streamlined manufacturing cycles and diminished lead instances.
Samsung additionally plans to broaden the scope of its technical partnership with key companions, based mostly on shut discussions with international GPU producers and hyperscalers targeted on next-generation ASIC improvement.
Samsung anticipates that its HBM gross sales will greater than triple in 2026 in comparison with 2025, and is proactively increasing its HBM4 manufacturing capability. Following the profitable introduction of HBM4 to market, sampling for HBM4E is anticipated to start within the second half of 2026, whereas customized HBM samples will begin reaching prospects in 2027,…
