Particle.news

Nvidia Pushes Memory Makers to Use TSMC 3nm for Custom NVHBM

Reports say the move is meant to cut per-vendor testing and ensure more consistent memory performance.

Overview

  • News reports say Nvidia has asked Samsung, Micron and SK hynix to build the NVHBM base die on TSMC’s 3nm process and that the three firms are cooperating on the effort, but manufacturers have not publicly confirmed the arrangement.
  • Samsung is reported to have agreed to use TSMC’s 3nm base die for NVHBM and to temporarily set aside its usual turnkey approach while it reviews its own foundry capacity.
  • Micron is reported to be outsourcing both HBM4E and NVHBM base dies to TSMC with volume production expected in 2027, and SK hynix is said to prioritize 3nm for future HBM releases.
  • Nvidia’s design would place its memory controller in the base die and target a sixth-generation (1c) DRAM core die, with company claims that NVHBM could deliver up to 30% more bandwidth and up to 15% lower power versus HBM4E; those performance goals remain unproven in shipping products.
  • Consolidating base-die work at TSMC would reduce the need to verify each vendor’s samples, tighten TSMC’s role in the AI memory supply chain, and raise questions about foundry capacity and next steps for memory makers and customers.