韓國記憶體廠商熱議 HBF 下一代突破

🔑 增強重點摘要

• •HBF employs die stacking, through-silicon vias (TSVs), and DDR synchronous signaling to achieve up to 800 GB/s aggregate bandwidth and 1.6 TiB capacity per stack[1].
• •HBF targets read-oriented workloads like LLM inference, offering 20-30% better energy efficiency at 0.48 nJ/byte with multi-way interleaving compared to conventional flash[1].
• •HBF maintains backward compatibility with existing flash pinouts and supports fallback to single-data-rate timing without new pins[1].
• •SanDisk has developed HBF to deliver performance within 2.2% of unlimited-capacity HBM while scaling capacity significantly for AI applications[8].

🛠️ 技術深入

• •Architecture: NAND flash with multi-layer die stacking (each die has hundreds of 3D NAND layers), TSVs for vertical interconnects, and HBM-style host interfaces[1][7].
• •Bandwidth: 400–800 GB/s aggregate per stack; single-channel DDR achieves 2–3× gains over SDR flash interfaces[1].
• •Capacity: ~1.6 TiB per stack, 16× larger than HBM4's ~100 GiB[1].
• •Energy: 0.48 nJ/byte at sixteen-way SLC interleaving, 20-30% reduction vs. conventional[1].
• •Interface: Multi-channel/way interleaving, DDR synchronous signaling, backward-compatible pinout and SDR fallback[1].

🔮 前景展望AI analysis grounded in cited sources

⏳ 時間線