Data
GPU Database
GPU database: NVIDIA H100, H200, B200 and GB200 plus AMD Instinct accelerators — specifications, use cases and where to rent them.
| GPU | Manufacturer | Architecture | Memory | Bandwidth | Availability |
|---|---|---|---|---|---|
| AMD Instinct MI300X | AMD | CDNA 3 | 192 GB HBM3 | Approx. 5.3 TB/s | Announced 2023 |
| AMD Instinct MI325X | AMD | CDNA 3 | 256 GB HBM3E | 6 TB/s | Introduced 2024 |
| AMD Instinct MI350X | AMD | CDNA 4 | 288 GB HBM3E | Up to 8 TB/s | Launched June 2025 |
| AMD Instinct MI355X | AMD | CDNA 4 | 288 GB HBM3E | Up to 8 TB/s | Launched 2025 |
| NVIDIA A100 | NVIDIA | Ampere | 80 GB HBM2e | Over 2 TB/s | Introduced 2020 |
| NVIDIA B200 | NVIDIA | Blackwell | 180 GB HBM3e | Up to 8 TB/s | Announced 2024 |
| NVIDIA B300 | NVIDIA | Blackwell Ultra | 288 GB HBM3e | Up to 8 TB/s | Shipping in NVIDIA DGX B300 systems in 2026 |
| NVIDIA GB200 | NVIDIA | Blackwell | 372 GB HBM3e per Grace Blackwell Superchip | 16 TB/s per Grace Blackwell Superchip | Announced 2024 |
| NVIDIA GH200 Grace Hopper Superchip | NVIDIA | Grace Hopper | 96 GB HBM3 or 144 GB HBM3e GPU memory | Up to 4.9 TB/s GPU memory bandwidth | Introduced 2023 |
| NVIDIA H100 | NVIDIA | Hopper | 80 GB HBM3 (SXM) / 80 GB HBM2e (PCIe) | Up to 3.35 TB/s | Announced 2022 |
| NVIDIA H200 | NVIDIA | Hopper | 141 GB HBM3e | Approx. 4.8 TB/s | Announced 2023 |
| NVIDIA L40S | NVIDIA | Ada Lovelace | 48 GB GDDR6 with ECC | 864 GB/s | Introduced 2023 |
Comparing data-center GPUs
The accelerators that train and serve large AI models differ mainly in memory capacity, memory bandwidth, interconnect and power draw. NVIDIA's Hopper generation (H100, H200) remains widely deployed; Blackwell (B200, and the GB200 superchip that pairs Blackwell GPUs with a Grace CPU) raises performance and memory further. AMD's Instinct line is the main merchant alternative.
Each GPU page lists specifications, typical use cases, published cloud prices and recent news about that chip.