NVIDIA B200 (Blackwell)
NVIDIA's newest datacenter GPU, the building block of the DGX B200 system below.
180 GB
GPU Memory vs. tier max
1000 W
Power vs. tier max
N/A
Price vs. tier max
| The dedicated memory built into the graphics card. An AI model's entire set of "weights" (its learned parameters) must fit into this memory before the card can run it. If a model needs more VRAM than one card has, you need multiple cards working together, or a card with more memory. | 180 GB |
|---|---|
| Thermal Design Power — roughly how many watts the card draws under full load. Higher-performance cards need more electricity and generate more heat, which affects your electricity bill and cooling needs. | 1000 W |
| Approximate cost to acquire the hardware. For consumer cards this is usually a fixed retail price. For datacenter-class hardware, list prices are rarely published — buyers negotiate directly with NVIDIA or its partners, so we show a typical range or say "contact for pricing" rather than inventing a number. |
No public price
No public price — contact NVIDIA/partners for a quote (industry estimates range $30,000-$50,000/chip)
|
Note on this spec
NVIDIA's own DGX B200 documentation implies 180GB usable memory per GPU (1,440GB / 8); some architecture-launch press cites 192GB physical HBM3e per GPU. We show NVIDIA's own figure and flag the discrepancy. The 1,000W figure is press-consensus, not an NVIDIA-published per-GPU number.Why this matters
A card or server needs at least as much
The dedicated memory built into the graphics card. An AI model's entire set of "weights" (its learned parameters) must fit into this memory before the card can run it. If a model needs more VRAM than one card has, you need multiple cards working together, or a card with more memory.
as an AI model requires,
or the model simply won't load. See Get a Recommendation
to see which models this hardware can run, or
add it to a side-by-side comparison.