Samsung has officially launched its HBM4 memory technology into commercial production, marking a pivotal moment in the AI hardware race. The new memory, based on a 4nm logic die and the company’s sixth-generation 1c DRAM, is now being deployed in NVIDIA’s next-generation Vera Rubin AI platform, setting a new benchmark for speed and capacity in high-bandwidth applications.
The HBM4 modules deliver pin speeds of up to 11.7 Gbps—nearly 50% faster than the previous 8 Gbps standard—with overclocking pushing performance further to 13 Gbps. This aligns with NVIDIA’s requirements for Vera Rubin, where memory bandwidth and latency are critical for AI inference and training workloads.
Samsung’s current HBM4 solutions focus on a 12-layer stack, with a 16-layer variant in development. The latter would enable per-module capacities of up to 48 GB, addressing the growing demand for dense, high-performance memory in AI and data center environments.
Key Upgrades Over Previous HBM Generations
- Speed: 11.7 Gbps (standard), 13 Gbps (overclocked) – a 46% jump over 8 Gbps HBM3.
- Process: 4nm logic die paired with 6th-gen 1c DRAM for efficiency.
- Capacity: 12-layer modules in production; 16-layer (48 GB) in development.
- Integration: First commercially deployed HBM4 solution, validated for NVIDIA’s Vera Rubin.
- Scalability: Designed to support future HBM4E variants (expected in H2 2026).
For AI workloads, this means reduced latency and higher throughput—critical for real-time processing in applications like generative AI, autonomous systems, and large-scale simulations. The 16-layer architecture, once mature, could further extend the reach of single-node AI systems by consolidating memory capacity without sacrificing speed.
Samsung’s push into HBM4 comes as the memory market consolidates among SK hynix, Micron, and itself. While SK hynix remains the dominant player, Samsung’s entry—backed by NVIDIA’s adoption—positions it as a key supplier for next-gen AI infrastructure. The company has already signaled aggressive growth, projecting a threefold increase in HBM revenue for 2026 compared to last year.
With Vera Rubin targeting enterprise and cloud deployments, Samsung’s HBM4 could become the default choice for NVIDIA’s most demanding customers, from hyperscalers to research labs. The technology’s scalability also hints at broader adoption in supercomputing and specialized AI accelerators.
Availability for HBM4 in Vera Rubin is already underway, though broader market rollout will depend on Samsung’s production scaling and NVIDIA’s roadmap timelines. For data center operators, the upgrade path is clear: HBM4 represents a tangible leap in performance-per-watt for AI-driven workloads.
