对比
芯片、系统与模型的并排规格对比。每项数据均保留其来源标注——厂商声明与估算值并列时，均会明确标注。
| NVIDIA H100 The GPU of the ChatGPT era — the unit of account for the 2023–25 AI build-out. | AMD Instinct MI300X AMD's serious challenge to Hopper — more memory than H100, adopted by Microsoft and Meta. | |
|---|---|---|
| Overview — Designer | NVIDIA | AMD |
| Overview — Architecture | Hopper (GH100) | CDNA 3, chiplet (8 XCD + 4 IOD) |
| Overview — Announced | March 2022 (GTC); shipping volume from early 2023 | Launched December 2023 |
| Overview — Status | Shipping — succeeded by Blackwell at the frontier | Shipping |
| Manufacturing — Foundry / process | TSMC 4N (custom 5 nm-class)Vendor spec | TSMC N5 compute dies on N6 base diesVendor spec |
| Manufacturing — Transistors | 80 billionVendor spec | 153 billionVendor spec |
| Manufacturing — Die size | 814 mm²Vendor spec | — |
| Manufacturing — Packaging | CoWoS-S with 5 or 6 HBM stacksVendor spec | CoWoS + SoIC 3D hybrid bondingVendor spec |
| Memory — Memory | 80 GB HBM3 (SXM); 94 GB HBM3 (NVL); 80 GB HBM2e (PCIe)Vendor spec | 192 GB HBM3 (8 stacks)Vendor spec |
| Memory — Memory bandwidth | 3.35 (SXM) / 3.9 (NVL) TB/sVendor spec | 5.3 TB/sVendor spec |
| Compute — TF32 tensor (dense) | 495 TFLOPSVendor spec | — |
| Compute — BF16/FP16 tensor (dense) | 990 TFLOPSVendor spec | — |
| Compute — FP8 tensor (dense) | 1,979 TFLOPSVendor spec | — |
| Compute — INT8 tensor (dense) | 1,979 TOPSVendor spec | — |
| Compute — FP64 tensor | 67 TFLOPSVendor spec | — |
| AI Performance — MLPerf | Set the reference results for training and inference benchmarks across 2023–24 submissionsMLPerf | — |
| Power — TDP | up to 700 (SXM) / 350 (PCIe) WVendor spec | 750 WVendor spec |
| Physical — Form factors | SXM5, PCIe, NVL (paired PCIe) | — |
| Physical — Interconnect | NVLink 4 — 900 GB/s; PCIe 5.0Vendor spec | Infinity Fabric — 896 GB/s aggregate per GPUVendor spec |
| Price — Launch price | Not publicly disclosed | — |
| Price — Market price | ~$25,000–40,000 per GPU at the 2023–24 peakReported | — |
| Price — Cloud rental | Roughly $2–7 per GPU-hour across major and specialist clouds during 2025 — falling steadilyReported | — |
| Compute — BF16/FP16 (dense) | — | 1,307 TFLOPSVendor spec |
| Compute — FP8 (dense) | — | 2,615 TFLOPSVendor spec |
| Compute — FP64 matrix | — | 163.4 TFLOPSVendor spec |
| Physical — Form factor | — | OAM; 8-GPU UBB platformVendor spec |
| Price — Price | — | Reported meaningfully below H100 street pricing; AMD does not publish list pricesReported |
"—"表示该记录未公布此项数据；不同厂商披露内容各异，我们不以猜测填补空白。