Tool · GPU & Accelerator Comparator
H100, B200, MI300X, TPU v5p, Trainium2 — side by side.
Every spec quoted here is from the official datasheet. Pick up to 4 accelerators to compare. Derived per-watt and per-GB ratios update automatically. Use this before you argue about benchmarks.
Select up to 4 accelerators
| Spec | H200 NVIDIA | B200 SXM NVIDIA | MI300X AMD | Trainium2 AWS |
|---|---|---|---|---|
| Released | 2024 | 2025 | 2023 | 2024 |
| Architecture | Hopper | Blackwell | CDNA 3 | Trainium2 |
| Process | TSMC 4N | TSMC 4NP (dual die) | TSMC 5nm + 6nm (chiplet) | TSMC 5nm |
| Die | 814 mm² | 2 × 814 mm² (NV-HBI linked) | 8× XCD chiplets + IOD | 2 compute tiles + 4 HBM stacks |
| Transistors | 80B | 208B total | 153B | (undisclosed) |
| Memory (GB) | 141 GB | 192 GB | 192 GB | 96 GB |
| Memory type | HBM3e | HBM3e (8 stacks) | HBM3 | HBM3 |
| Memory BW | 4.80 TB/s | 8.00 TB/s | 5.30 TB/s | 2.90 TB/s |
| FP64 (TFLOPS) | 67 TFLOPS | 40 TFLOPS | 163 TFLOPS | — |
| BF16 | 0.99 PFLOPS | 2.25 PFLOPS | 1.30 PFLOPS | 0.65 PFLOPS |
| FP8 | 1.98 PFLOPS | 4.50 PFLOPS | 2.60 PFLOPS | 1.30 PFLOPS |
| FP4 | — | 9.00 PFLOPS | — | — |
| TDP | 700 W | 1000 W | 750 W | 500 W |
| Scale-up | NVLink 4 (900 GB/s) | NVLink 5 (1.8 TB/s) | Infinity Fabric (896 GB/s, 8-way) | NeuronLink (64-chip UltraServer) |
| Scale-up BW | 0.90 TB/s | 1.80 TB/s | 0.90 TB/s | 0.50 TB/s |
| Scale-out | NDR InfiniBand 400G | NDR/XDR InfiniBand | RoCE / InfiniBand | EFA (Elastic Fabric Adapter) |
| Approx price | ~$32k | ~$38k | ~$18k | — |
| Availability | Shipping since mid-2024 | Ramping 2025-2026 | Shipping since late 2023 | AWS only |
| Notes | Same silicon as H100, HBM3e refresh: 76% more memory + 43% more bandwidth. | First dual-die AI GPU, connected via NV-HBI (10 TB/s). Liquid cooling required. | Largest HBM capacity of any 2023-era GPU. ROCm software stack closed the gap but still behind CUDA. | Powers Project Rainier (500K+ chips for Anthropic). Only via AWS EC2 Trn2 instances. |
📊 Derived ratios (per chip)
| Metric | H200 | B200 SXM | MI300X | Trainium2 |
|---|---|---|---|---|
| Memory per watt (GB/W) | 0.201 | 0.192 | 0.256 | 0.192 |
| BF16 PFLOPS per watt | 1.41 | 2.25 | 1.73 | 1.30 |
| Memory BW per watt (GB/s/W) | 6.86 | 8.00 | 7.07 | 5.80 |
| $/GB HBM (at list) | $227/GB | $198/GB | $94/GB | — |
Specs from NVIDIA H100/H200/B200/GB200 datasheets, AMD Instinct MI300X/MI325X sheets, Google Cloud TPU v5p docs, AWS Trainium2 announcement, Cerebras WSE-3 specs. Street prices are approximate and for orientation only — actual pricing varies by volume + contract. Highlighted cells are best-in-comparison; exclusions like GB200 NVL72 (a rack, not a chip) will skew ratios.