Samsung Completes HBM4 Logic Die Design with 4nm Process

According to reports from Chosun Ilbo, Samsung's Device Solutions (DS) division has successfully completed the design of its HBM4 (High-Bandwidth Memory) logic chip, marking a significant step forward in memory technology. Samsung Foundry has already begun pilot production using a 4nm process based on this design. After completing final performance validation of the Logic Base Die, Samsung is preparing to offer HBM4 samples for further verification.

Unlike previous generations of HBM, such as HBM3E, which used DRAM-based Base Dies, HBM4 integrates a Logic Base Die to enhance both performance and energy efficiency. The Logic Base Die serves as a crucial component that links AI accelerators' GPUs with DRAM, located beneath the DRAM stack. It acts as a controller, facilitating communication between the GPU and memory. This innovation allows for greater customization, as clients can integrate their own intellectual property (IP) into the design, improving data processing efficiency. Furthermore, the new configuration is expected to reduce power consumption by up to 70% compared to previous generations.

The Logic Base Die in HBM4 is akin to the "brain" of the chip, responsible for controlling the stacked DRAM layers above. The increased number of memory I/O pins and the need for additional integrated functions in HBM4 are driving the adoption of logic semiconductor foundry technologies for its production.

Experts note that heat generation is a primary concern for HBM, especially during operation, with the Logic Base Die being a major heat contributor. Advanced manufacturing processes are crucial for mitigating heat and improving overall efficiency and performance.

Samsung aims to regain market share lost with the HBM3E by pursuing an aggressive strategy with HBM4. In addition to using its own 4nm logic-based Base Die, Samsung plans to introduce 10nm-level DRAM production and hopes to achieve 16-layer stacking with bump-free hybrid bonding for enhanced performance.

There are also reports that Samsung may provide customized HBM4 memory for major AI cloud service providers Meta and Microsoft, integrating into their next-generation AI solutions. Samsung's customized HBM4 is expected to be used in Microsoft's AI chip "Maia 100," and Meta's Artemis AI chip may also incorporate this advanced memory. This development is seen as a promising step for Samsung as it strives to capture a larger share of the growing AI market, narrowing the gap with competitors like SK Hynix.