As MediaTek officially joins Google's eighth-generation TPU project, its role in the artificial intelligence ecosystem is constantly strengthening, further consolidating its technology accumulation and market influence in the field of customized chip solutions.MediaTek CEO Cai Lixing recently systematically elaborated on the four major technical challenges faced by XPU development at a public event, covering computing power, memory bottlenecks, interconnect efficiency, and advanced packaging technology.

in,Memory has become a key variable affecting system performance and cost structure. It currently accounts for up to 50% of the XPU bill of materials, highlighting the decisive role of memory in the cost and performance of the overall solution.

Cai Lixing pointed out that although AI training tasks still mainly rely on HBM (high-bandwidth memory), as market demand gradually evolves towards customization and high-performance reasoning, AI reasoning is becoming the next important growth engine.

Under this trend, DDR DRAM is expected to be more widely used in inference scenarios due to its higher density and cost-effectiveness, while SRAM will be reserved for specific selective scenarios. This has also led memory giants such as SK Hynix and Samsung to accelerate the layout of related technologies.

Among them, SK hynix has launched a multi-dimensional layout around the "AI-N" series product line, with "AI-N P" (performance), "AI-N B" (bandwidth) and "AI-N D" (density) as the three major technical directions.They correspond to SLC NAND flash memory solutions, high-bandwidth flash memory (HBF) in cooperation with NVIDIA, and data center solutions for large-capacity and low-power consumption requirements, striving to provide differentiated technical support in the context of the growing AI inference load.

Samsung continues its layout in the field of in-memory computing. As early as 2021, Samsung launched the industry's first HBM with integrated AI processing capabilities - HBM-PIM, which provides up to 1.2 TFLOPS of embedded computing power, allowing the memory chip itself to perform some tasks of the CPU, GPU, ASIC or FPGA. Recent news has pointed out that Samsung is refocusing on PIM technology research and development, with the intention of promoting it to gradually replace the traditional HBM architecture in future AI applications.