Detailed stacking technology, bandwidth throughput, and NVIDIA Blackwell allocation breakdown across SK Hynix MR-MUF and Samsung NCF.
| Generation | Manufacturer | Bandwidth | Market Share | Stack Technology | Primary Clients |
|---|---|---|---|---|---|
| HBM3e (36GB 12-Hi) | SK Hynix | 1.28 TB/s | 85% | Advanced MR-MUF (Mass Reflow Molded Underfill) | NVIDIA GB200, B200, H200 |
| HBM4 (48GB 16-Hi) | SK Hynix | 2 TB/s | 70% | Custom TSMC Base Die + MR-MUF | NVIDIA Rubin R100, AMD MI400 |
| HBM3e (36GB 12-Hi Shinebolt) | Samsung Electronics | 1.28 TB/s | 15% | Advanced NCF (Non-Conductive Film) | AMD MI325X, CSP Custom ASICs |
| CXL 2.0 Memory Expander (CMM-D) | Samsung Electronics | 0.51 TB/s | 60% | PCIe 5.0 CXL Controller | Hyperscale Cloud Host Nodes |
| HBM4 (48GB 16-Hi Turnkey) | Samsung Electronics | 2 TB/s | 30% | Custom 4nm GAA Base Die + NCF | Major Hyperscalers & ASICs |
| LPDDR5X CAMM2 Modular AI Memory | SK Hynix | 0.38 TB/s | 65% | Low-Power Modular Architecture | AI PC OEMs & Edge Accelerators |