Micron said in its latest quarterly results that it will work with TSMC to produce base logic dies for both standard and custom HBM4E memory. The disclosure highlights a significant shift in how next-generation high-bandwidth memory may be built, as foundries, rather than DRAM makers, begin handling the foundational logic layer of advanced HBM stacks.
TSMC roadmap revealed
According to German outlet Hardwareluxx, TSMC told attendees at a recent forum in Amsterdam that customized HBM will become a formal product class in the HBM4E generation, referred to as C-HBM4E.
The report stated that beginning with HBM4 and C-HBM4E, base dies will reportedly no longer be produced by memory suppliers such as Samsung Electronics, SK Hynix, and Micron. Instead, standardized base dies will reportedly be manufactured by foundries using advanced logic nodes and designed for direct integration into multi-die packages.
Shift to advanced nodes
Hardwareluxx reported that HBM4 base dies will be manufactured on TSMC's N12 process, with operating voltage reduced from 1.1 volts to 0.8 volts, which could raise efficiency by about 1.5 times.
For C-HBM4E, the base die will reportedly move to the N3P node, with voltage lowered further to 0.75 volts. Efficiency could reach roughly twice that of HBM3E, according to the report.
CoWoS-L advances
TSMC is advancing its CoWoS-L packaging platform to support the growing power and memory requirements of AI accelerators. Hardwareluxx reported that the next version of CoWoS-L will support a mask size limit of 5.5 times and integrate up to 12 HBM3E or HBM4 stacks. The design is reportedly aimed at products such as AMD's Instinct MI450X and Nvidia's Vera Rubin accelerators expected in 2026.
A follow-on CoWoS-L generation planned for 2027 will reportedly adopt the A16 process and expand mask size limits to 9.5 times, enabling system designs that incorporate more than 12 HBM stacks.
Broader industry impact
Industry observers cited by Hardwareluxx noted that chip designers in high-performance computing and AI are unlikely to avoid TSMC in the coming years. The company integrates manufacturing and electronic design automation tools in a way that allows deep involvement early in the chip development cycle.
Beyond its advanced logic nodes, TSMC has become a central provider of multi-chip packaging technologies required for complex HBM-based architectures. These combined capabilities are expected to strengthen its position as a key supplier for next-generation AI and HPC products.
Article translated by Sherri Wang and edited by Jack Wu
