Samsung Raises the Bar Again: First Samples of HBM4E Memory Shipped. The Features

Samsung has announced the start of shipments to major global clients of the first samples of HBM4E memory with 12 layers, a new evolution in its roadmap for high-bandwidth memory dedicated to AI workloads and hyperscale infrastructures.

The announcement comes just a few months after the start of mass production of HBM4, a technology with which Samsung had already claimed industrial leadership in the segment. With HBM4E, the company aims to further increase bandwidth, capacity, and energy efficiency, aspects that are increasingly central in the training and inference of large-scale AI models.

According to the Korean company, the new HBM4E reaches a stable speed of 14 Gbps per pin, with the possibility of scaling up to 16 Gbps. Compared to the previous HBM4, the improvement exceeds 20%, while the bandwidth per stack reaches up to 3.6 TB/s, a value designed to meet the computational needs of LLMs and next-generation AI accelerators.

The initial configuration proposed by Samsung features a 12-layer package with a capacity of 48 GB, over 30% more than the previous generation. The company also announced its intention to introduce 32 GB variants with 8-layer stacks and 64 GB models with 16-layer configurations, dependent on customer requests.

From a production standpoint, Samsung emphasizes that it has reused part of the experience gained with HBM4. The HBM4E utilize DRAM based on the 1c process at 10 nanometers from the sixth generation, paired with a logic base die produced at Samsung Foundry using 4 nm technology.

According to the company, the joint optimization of the logic architecture and memory has allowed for improvements not only in performance but also in yield and energy efficiency. Samsung reports a 16% increase in efficiency compared to the previous generation and a thermal resistance reduction of over 14%, a particularly important factor in high-density AI data centers.

Improvements in terms of thermal management and power consumption are becoming increasingly relevant in the AI market, where accelerators and GPUs operate with high energy consumption and continuous loads. More efficient dissipation indeed allows for sustained high performance while simultaneously reducing the energy costs of infrastructures.

Samsung has confirmed that the mass production of HBM4E will begin following the timelines agreed upon with commercial partners, after the initial phase of validation and optimization of the samples currently being distributed. The company also highlights that the feedback received from customers regarding the HBM4 introduced in February has been particularly positive, especially in terms of performance and energy efficiency. The HBM4 had reached speeds of 11.7 Gbps.