HBM4 2048-Bit Memory & TSMC Logic Dies: Will SK Hynix Keep Its AI Lead Over Samsung?
Doubling High Bandwidth Memory interfaces from 1,024 bits to 2,048 bits in HBM4 has forced memory makers SK Hynix, Samsung, and Micron to replace conventional DRAM buffer bases with custom foundry logic base dies fabricated on TSMC 3-nanometer and 12-nanometer nodes for Nvidia's Rubin architecture.
Through the HBM3E generation powering Nvidia's Blackwell GPUs, memory manufacturers fabricated the bottom base die using their own internal DRAM semiconductor processes. However, routing 2,048 vertical data channels per stack—doubling the pin count to smash the memory bandwidth wall—exceeds the wiring density and thermal limits of legacy DRAM lithography. By migrating the base die to TSMC's N12 and N3P FinFET logic nodes (or Samsung Foundry's 4nm process), AI chipmakers can offload portions of the physical memory controller and custom compute logic directly into the bottom of the HBM4 stack, freeing up to 25% of die area on the main GPU reticle.
This architectural convergence is reshaping the semiconductor supply chain as 12-high and 16-high HBM4 stacks transition from traditional solder micro-bumps and Mass Reflow Molded Underfill (MR-MUF) toward direct copper-to-copper hybrid bonding. Because hyperscalers like Google, AWS, and Meta are demanding custom HBM4 (cHBM) tailored to their proprietary AI accelerators, memory suppliers can no longer sell commoditized JEDEC stacks without deep foundry co-design.
Will custom HBM4 logic base dies cement TSMC and SK Hynix's packaging dominance, or can Samsung's vertically integrated memory-and-foundry model capture Nvidia's Rubin Ultra orders?
Comentarios
Sé respetuoso. Los comentarios que reporten varias personas se ocultan hasta que los revise un moderador.
Para leer no necesitas cuenta; solo para comentar. Entras con un código que te llega al correo o con tu cuenta de Google. Mostramos solo el nombre que elijas; nunca tu correo.