Tech news in 3 minutes
Containing AI Agents With a Will of Their Own
SK hynix’s Hot Chips 2026 presentation reveals that High Bandwidth Memory (HBM) evolution for AI now depends on advanced packaging, thermal management, and processor-memory co-design, not just DRAM circuit improvements. The memory giant compared four HBM3E packages with twelve GDDR6 devices, showing HBM delivers 144GB capacity and ~4 TB/s bandwidth while using half the area and substantially lower energy per bit. Each HBM generation advances bandwidth, capacity, and power efficiency simultaneously: HBM2E at ~460 GB/s, HBM3 at 717 GB/s, HBM3E at 1,024 GB/s, and HBM4 targeting 2,048 GB/s by doubling interface width to 2,048 data connections. HBM4 contains over 20,000 through-silicon vias (TSVs) and ~16,148 micro-bumps, creating manufacturing challenges in wafer thinning, die warpage, bump uniformity, and heat removal. SK hynix highlights mass reflow with molded underfill (MR-MUF) as a key production technology, enabling a 16-layer HBM3E stack with 48GB within 775µm height. For stacks of 20+ dies, hybrid bonding—joining dielectric surfaces and copper directly without solder bumps—is identified as the successor, allowing narrower connections, more TSVs, and thicker DRAM dies. Thermal management is critical: the proposed i-HBM architecture inserts a thermally conductive insulator into die-to-die interfaces, achieving over 30% reduction in thermal resistance. The presentation underscores a broader industry shift: memory is now integrated early alongside processors and interposers in advanced AI packages, requiring collaboration among memory suppliers, foundries, processor designers, and packaging companies. Advanced packaging is no longer an assembly detail but the determinant of future AI systems’ bandwidth, capacity, reliability, and energy efficiency.