Core Viewpoint - SK Hynix has introduced "StreamDQ" as a solution to address the upcoming era of customized High Bandwidth Memory (HBM), enhancing data processing performance by offloading certain tasks from GPUs to HBM [1][2]. Group 1: StreamDQ Technology - StreamDQ technology aims to commercialize customized HBM, with a focus on promoting it to major clients like NVIDIA at CES 2026 [2]. - The technology transfers some controller functions from existing GPUs to the HBM substrate, potentially improving the performance and efficiency of system semiconductors [2]. - StreamDQ allows real-time dequantization of data as it flows through HBM, which can significantly reduce the bottleneck traditionally faced by GPUs during the dequantization process [3]. Group 2: Performance Enhancements - The implementation of StreamDQ is expected to enhance the inference speed of large language models (LLMs) by approximately seven times, addressing the memory bottleneck that previously consumed up to 80% of the overall inference time [3]. - By processing data closer to memory, the overall efficiency of semiconductor systems is anticipated to improve significantly, aligning with the concept of processing near memory (PNM) [3]. Group 3: Manufacturing and Integration - SK Hynix is leveraging advanced processes from Taiwan's TSMC for adding GPU controllers and other components to the base chip without facing major challenges [2]. - The integration of UCIe interfaces on the chip substrate enhances chip integration, allowing for the segmentation of chips into functional units for manufacturing [2].