Scaling the Memory Wall: The Rise and Roadmap of HBM
SemiAnalysis 1 year ago 25
High Bandwidth Memory (HBM) has become essential for AI accelerators, with demand driven by AI model scaling and accelerator designs from Nvidia, Google, OpenAI, and others. The report projects Nvidia will command the largest share of HBM bit demand in 2027, with its Rubin Ultra pushing per-GPU capacity to 1 TB, while demand grows across custom ASICs and hyperscalers like Amazon and Google. Manufacturing challenges including power delivery network design, thermal management, and bonding tool precision create yield pressures that differentiate suppliers, with SK Hynix and Micron achieving higher yields than Samsung, thereby sustaining premium HBM pricing and margin growth.