Compare
Side-by-side spec sheets for chips, systems and models. Every figure keeps its basis label — a vendor claim next to an estimate is labeled as exactly that.
| AWS Trainium2 The chip under Project Rainier — Anthropic's half-million-chip training fleet. | NVIDIA H100 The GPU of the ChatGPT era — the unit of account for the 2023–25 AI build-out. | |
|---|---|---|
| Overview — Designer | AWS (Annapurna Labs) | NVIDIA |
| Overview — GA | December 2024 | — |
| Overview — Status | Shipping | Shipping — succeeded by Blackwell at the frontier |
| Memory — Memory | 96 GB HBM3Vendor spec | 80 GB HBM3 (SXM); 94 GB HBM3 (NVL); 80 GB HBM2e (PCIe)Vendor spec |
| Memory — Memory bandwidth | 2.9 TB/sVendor spec | 3.35 (SXM) / 3.9 (NVL) TB/sVendor spec |
| Compute — FP8 (dense) | 1.3 PFLOPSVendor spec | — |
| Compute — Sparse FP8 | 5.2 PFLOPS (16:4 sparsity)Vendor spec | — |
| Physical — UltraServer | 64 chips, 83.2 FP8 PFLOPS dense, 6 TB HBMVendor spec | — |
| Physical — Interconnect | NeuronLink; EFA scale-out networkingVendor spec | NVLink 4 — 900 GB/s; PCIe 5.0Vendor spec |
| Overview — Architecture | — | Hopper (GH100) |
| Overview — Announced | — | March 2022 (GTC); shipping volume from early 2023 |
| Manufacturing — Foundry / process | — | TSMC 4N (custom 5 nm-class)Vendor spec |
| Manufacturing — Transistors | — | 80 billionVendor spec |
| Manufacturing — Die size | — | 814 mm²Vendor spec |
| Manufacturing — Packaging | — | CoWoS-S with 5 or 6 HBM stacksVendor 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 |
| Physical — Form factors | — | SXM5, PCIe, NVL (paired PCIe) |
| 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 |
"—" means that record does not publish this line; different vendors disclose different things, and we do not fill gaps with guesses.