Every GPU in Our Directory, Ranked by VRAM
Memory is the first question in local AI. A model either fits in your card's VRAM or it does not, and no driver setting changes that. So we lined up all 171 GPUs in this directory by the one number that decides fit: VRAM, from the 288 GB data center accelerators at the top to the 8 GB entry cards at the bottom.
The ladder splits into clear bands. Data center parts (B300, MI350X, H200 class) hold 128 to 288 GB and run the biggest open models whole. Workstation cards top out at 96 GB with the RTX PRO 6000 Blackwell. Consumer cards run from 32 GB on the RTX 5090 down to 8 GB, and that spread is where most home buying decisions live: 16 GB is the floor for comfortable 8B work at FP16, 24 GB is the sweet spot for 32B models at INT4 and 70B models with offload, and 32 GB is where a single card starts to feel roomy.
The top 25 GPUs by VRAM
| GPU | VRAM | AI TOPS | TDP | US list price |
|---|---|---|---|---|
| NVIDIA B300 SXM | 288 GB | Not published | 1400 W | $69,999 |
| NVIDIA GB300 NVL72 | 288 GB | Not published | 1400 W | $3,750,000 |
| AMD Instinct MI350X | 288 GB | Not published | 1400 W | $39,999 |
| AMD Instinct MI355X | 288 GB | Not published | 1400 W | $42,999 |
| AMD Instinct MI325X | 256 GB | Not published | 1000 W | $34,999 |
| NVIDIA B200 SXM | 192 GB | Not published | 1200 W | $54,999 |
| NVIDIA GB200 NVL72 | 192 GB | Not published | 1200 W | $3,000,000 |
| AMD Instinct MI300X | 192 GB | Not published | 750 W | $29,999 |
| NVIDIA H200 SXM | 141 GB | Not published | 700 W | $42,999 |
| NVIDIA H200 NVL | 141 GB | Not published | 600 W | $39,999 |
| AMD Instinct MI300A | 128 GB | Not published | 760 W | $24,999 |
| AMD Instinct MI250X | 128 GB | Not published | 560 W | $8,999 |
| Intel Gaudi 3 AI Accelerator | 128 GB | Not published | 900 W | $14,999 |
| NVIDIA RTX PRO 6000 Blackwell Workstation Edition | 96 GB | 4000 TOPS | 600 W | $17,690 |
| NVIDIA GH200 Grace Hopper Superchip | 96 GB | Not published | 900 W | $44,999 |
| NVIDIA RTX PRO 6000 Blackwell Server Edition | 96 GB | 4000 TOPS | 600 W | $18,999 |
| Intel Gaudi 2 AI Accelerator | 96 GB | Not published | 600 W | $8,999 |
| NVIDIA H100 NVL | 94 GB | Not published | 400 W | $31,999 |
| NVIDIA H100 SXM | 80 GB | Not published | 700 W | $34,999 |
| NVIDIA H100 PCIe | 80 GB | Not published | 350 W | $27,999 |
| NVIDIA A100 80GB SXM | 80 GB | Not published | 400 W | $12,500 |
| NVIDIA A100 80GB PCIe | 80 GB | Not published | 300 W | $13,100 |
| AMD Instinct MI210 | 64 GB | Not published | 300 W | $5,499 |
| NVIDIA RTX 6000 Ada Generation | 48 GB | 1458 TOPS | 300 W | $6,800 |
| NVIDIA RTX PRO 5000 Blackwell | 48 GB | 2988 TOPS | 300 W | $4,750 |
How the directory spreads across the bands
Two things stand out. First, TOPS figures are published for only part of the ladder, mostly workstation and consumer cards; the data center rows rarely publish a comparable number, which is why the table shows the gap instead of filling it. Second, price and VRAM stop tracking each other above 48 GB: those parts are sold on quote, and the honest price column is the blank one.
If you are shopping, start from the model, not the card. Work out the memory your model needs at the precision you will actually run, using the VRAM Fit Checker, then read the ladder at that line and everything above it. The full filterable list lives on the GPU category page.