L4 · Chips
Data-Center GPU Comparison
The merchant data-center accelerators behind AI training and inference — NVIDIA Hopper/Blackwell vs AMD CDNA — compared on memory, bandwidth, and power. The interesting deltas are increasingly about HBM capacity and bandwidth, not raw FLOPs.
| Vendor | Accelerator | Architecture | Memory | Bandwidth | TDP | Year |
|---|---|---|---|---|---|---|
| NVIDIA | H100 (SXM) | Hopper | 80 GB HBM3 | 3.35 TB/s | 700 W | 2022 |
| NVIDIA | H200 (SXM) | Hopper | 141 GB HBM3e | 4.8 TB/s | 700 W | 2024 |
| NVIDIA | B200 | Blackwell | 192 GB HBM3e | 8 TB/s | 1000 W | 2024–25 |
| NVIDIA | GB200 (superchip) | Grace + 2× Blackwell | 384 GB HBM3e | 16 TB/s | ~2700 W | 2025 |
| AMD | MI300X | CDNA 3 | 192 GB HBM3 | 5.3 TB/s | 750 W | 2023 |
| AMD | MI325X | CDNA 3 | 256 GB HBM3e | 6.0 TB/s | 1000 W | 2024–25 |
| AMD | MI355X | CDNA 4 | 288 GB HBM3e | 8 TB/s | ~1400 W | 2025 |
Specs consolidated from NVIDIA and AMD datasheets. Per-SKU figures vary by form factor (SXM / OAM / PCIe).
The merchant GPU is only half the story
Hyperscaler custom ASICs now grow faster than merchant GPUs by unit volume, and HBM + advanced packaging gate how many of any of these actually ship. The sourced landscapes:
Engineering reference, not investment advice.