 |
Photo = Yonhap news |
[Alpha Biz= Reporter Kim Jisun] Samsung Electronics has started developing custom high-bandwidth memory (HBM4) to be supplied specifically to Microsoft and Meta.
According to the semiconductor industry on the 11th, Samsung is currently working on HBM4 for both companies.
Most major tech firms operate their own AI data centers, creating a strong need to reduce chip procurement costs. This drives them to purchase AI chips from companies like NVIDIA or AMD while also designing their own AI accelerator chips.
Samsung plans to complete the HBM4 development by the end of next year and move directly into mass production. Although specific details for the custom HBM4 to be supplied to these companies have not been disclosed, Samsung shared HBM4 specifications at the semiconductor conference ISSCC 2024 in February.
The data transfer speed, or bandwidth, is 2 terabytes (TB) per second, a 66% increase over HBM3E, and the DRAM stacking has been expanded to 16 layers, boosting capacity by 33% to 48 gigabytes (GB) from the current 36GB.
A semiconductor industry insider noted, "While heat management and speed have been key up to HBM3, HBM4 marks a shift where the ability to meet customer demands for AI processing (NPU) or specific data handling functions (IP) becomes crucial." Even with 16-layer HBM stacking, the ability to design and insert a custom "buffer die" as requested by clients is essential. According to TrendForce, the global HBM market is expected to reach $18.2 billion this year and expand by 2.5 times to $46.7 billion next year. With NVIDIA planning to mass-produce its next-generation GPU, "Rubin," in 2026, competition over HBM adoption is likely to intensify.
Alphabiz Reporter Kim Jisun(stockmk2020@alphabiz.co.kr)
