Samsung HBM3E 12-High 36GB Memory

The Samsung HBM3E 12-High 36GB is a High Bandwidth Memory third-generation extended specification module built on a 12-layer DRAM die stack, representing the highest-capacity single HBM3E package Samsung has produced as of 2025. Each package integrates 36GB of capacity through 12 stacked dies interconnected via through-silicon vias (TSVs) and micro-bumps, delivering aggregate memory bandwidth exceeding 1.2TB/s per stack. The architecture leverages Samsung's advanced 1c-nm DRAM process node to achieve an optimal balance of density, power efficiency, and signal integrity, making it the preferred memory solution for next-generation AI training accelerators, HPC processors, and custom AI ASICs where memory bandwidth is the primary system bottleneck.

HBM3E 12-High targets AI infrastructure deployments where the transition from HBM3 to HBM3E provides measurable gains in large language model (LLM) training throughput, multi-modal model inference, and scientific simulation workloads. The increased per-stack capacity of 36GB — up from the 24GB available in 8-high HBM3E configurations — allows AI accelerator designers to reduce the total package count on a substrate while maintaining or increasing total system memory capacity, directly benefiting thermal design and PCB routing density. This makes it especially valuable for chiplet-based accelerator designs integrating multiple compute dies with a shared HBM stack in a 2.5D or 3D package architecture.

Samsung engineered the HBM3E 12-High with enhanced ECC schemes across all 1024-bit wide bus channels to maintain data integrity under sustained high-bandwidth access patterns typical in transformer model matrix multiplications. The product complies with JEDEC HBM3E specifications and is compatible with advanced packaging platforms including CoWoS, Foveros, and SoIC, ensuring broad ecosystem integration with leading AI accelerator silicon from hyperscaler and merchant silicon vendors alike. Launched in Q2 2025, the 12-High variant positions Samsung at the forefront of the HBM supply chain for the next wave of 600W-class AI accelerator products.

Ideal for

• Large language model (LLM) pre-training on GPU and AI ASIC clusters where aggregate HBM bandwidth directly determines tokens-per-second throughput
• High-performance inference serving for multi-modal foundation models requiring low-latency, high-capacity on-package memory to minimize memory-bound bottlenecks
• Scientific HPC simulations including molecular dynamics, computational fluid dynamics, and climate modeling on memory-bandwidth-constrained accelerators
• Custom AI accelerator ASIC design for hyperscale data centers integrating HBM3E in 2.5D CoWoS packages to maximize flops-per-watt at the system level
• Sparse and dense recommendation model training at scale where large embedding tables benefit from the 36GB per-stack capacity reducing cross-package data movement
• Graph neural network (GNN) workloads with irregular memory access patterns that demand peak sustained bandwidth to maintain accelerator compute utilization

Technical specifications

Available from Omnixon Global. Submit an RFQ and our team will confirm configuration and availability for your order.