NVIDIA B200
18,432
CUDA Cores
192GB
VRAM
8000
GB/s
Technical Specifications
18,432
CUDA Cores
0
Base MHz
0
Boost MHz
192GB HBM3e
8192-bit bus
Performance
225
FP32 TFLOPS
450
FP16 TFLOPS
1000W
TDP
Cloud Availability
2
Available Instances
$3.50/hr
Starting Price
Detailed Specifications
| Architecture | Blackwell (Unknown) |
| Release Date | 2024-03-01 |
| Launch Price | $2,000.00 |
| Process | 4nm |
| Transistors |
AI Features
Gen 5
Tensor Cores
Enabled
Transformer Engine
Not Supported
Flash Attention
Physical Specifications
Dimensions
10.5in
Length
4.4in
Width
2-slot
Height
About B200 GPU
The NVIDIA B200 was introduced in 2024 on the Blackwell architecture. Its specification combines 18,432 CUDA cores with 192GB of HBM3e, which defines both the workloads it can run and the batch sizes it can hold.
Based on its 192GB memory capacity, the B200 is best evaluated for large-model training, high-batch inference, and memory-intensive scientific computing. Always size the model, optimizer state, KV cache, and framework overhead together rather than choosing a GPU from core count alone.
GPUvec currently tracks 2 cloud listings for the B200, with the lowest observed hourly rate at $3.50. Prices and availability can change by provider, region, and instance configuration, so verify the final rate before starting a workload.
For software planning, the B200 combines 8,000 GB/s of memory bandwidth, CUDA compute capability 10.0, and a 4nm manufacturing process. Confirm the minimum CUDA, PyTorch, TensorFlow, or driver version required by your application before renting or purchasing hardware.
Rent B200 from Our Partners
Get started quickly with these trusted GPU cloud providers. We may earn a commission when you sign up.
External Resources
Learn more about GPUs from these authoritative sources:
Official CUDA programming guide
NVIDIA GPU Specifications →Official NVIDIA GPU specs
TechPowerUp GPU Database →Comprehensive GPU specifications
CUDA Compute Capability Guide →GPU compute capability reference
What You Need to Know About the B200
Complete Specifications for the NVIDIA B200
The NVIDIA B200 uses the Blackwell architecture and combines 18,432 CUDA cores with 192GB of HBM3e. GPUvec records 8,000 GB/s memory bandwidth alongside architecture, release date, process node, accelerator features, and physical specifications so the page can be evaluated from source data instead of model-name assumptions.
Compare NVIDIA B200 Cloud Rental Prices per Hour
GPUvec currently tracks 2 cloud listings, with an observed starting rate of $3.50 per hour. Provider rates can vary by region, host configuration, billing model, and availability. Use the provider directory and related-GPU links to compare an available alternative when this exact accelerator is not listed.
Is the NVIDIA B200 the Right GPU for Your AI Workload?
Start with the 192GB memory limit, then evaluate 8,000 GB/s memory bandwidth, compute features, and software compatibility for your workload. Model weights, KV cache, activations, optimizer state, and framework overhead all consume memory. Compare this page with related accelerators rather than assuming that a higher core count alone guarantees a better training or inference result.
Top GPUs for Training and Inference
| Category | Rank 1 | Rank 2 | Rank 3 |
|---|---|---|---|
| Best for Training | NVIDIA H200 | NVIDIA H100 | NVIDIA B200 |
| Best for Inference | NVIDIA A40 | NVIDIA A100 | NVIDIA A10 |
Compare GPU specifications and cloud instances to find the best GPU for your workload.