Samsung bets on zHBM as 2.5D memory nears its limits

At SEMICON Taiwan 2026 Samsung presented a shift from 2.5D to 3D memory integration via vertically stacked zHBM. The goal is to raise AI response throughput from about 100 to 1000 tokens per second per user.

• zHBM stacks memory vertically instead of 2.5D layout
• Goal: AI response speed up from 100 to 1000 tokens/s
• Technology presented at SEMICON Taiwan 2026