180 GB HBM3e memory
Keep model weights, KV cache and working data in GPU memory. Context length, concurrency and training state determine the remaining capacity required.
Blackwell architecture with 180 GB of HBM3e per GPU. For large language model inference, fine-tuning and distributed training on compatible HGX systems.
Check the console or contact sales for prices and available instance configurations.
B200 combines high-capacity HBM3e, second-generation Transformer Engine and fifth-generation NVLink. Size the system for model state, numerical precision and GPU communication.
Keep model weights, KV cache and working data in GPU memory. Context length, concurrency and training state determine the remaining capacity required.
Second-generation Transformer Engine supports low-precision training and inference. FP4 and FP8 use requires compatible models, quantization, frameworks and kernels.
HGX B200 connects eight GPUs through NVLink and NVSwitch, with 1.8 TB/s of aggregate bidirectional bandwidth per GPU. Cross-node networking is a separate configuration.
Blackwell connects two dies through a 10 TB/s chip-to-chip interface and presents them as one CUDA GPU. This internal link is separate from NVLink between GPUs.
From model serving to distributed training, select resources for the workload's memory, compute precision and GPU communication needs.
Evaluate quantized LLM or MoE serving against the 180 GB HBM3e budget for weights, KV cache and runtime buffers. Select FP4 or FP8 only after checking kernel support and model accuracy.
For fine-tuning and other post-training jobs, budget activations, gradients and optimizer state alongside weights. Choose GPU count and parallelism from the complete training state, with storage planned for checkpoints.
When a model spans GPUs, evaluate the eight-GPU HGX topology and 1.8 TB/s bidirectional NVLink per GPU. Extending the job across nodes also requires a separately confirmed cluster network and communication strategy.
For text, image or video models, include encoder activations, input resolution and sequence length in the 180 GB memory budget. Verify operator and precision support before expanding image size, frame count or batch size.
The figures below are per GPU, not aggregate specifications for a complete multi-GPU system. Theoretical peaks differ from measured model performance and do not describe a complete instance's CPU, system RAM, storage or cross-node networking.
| Specification | Value per GPU |
|---|---|
| Architecture | BlackwellBlackwell GPU architecture. |
| GPU memory | 180 GB HBM3eStores model weights, KV cache and runtime state. |
| Memory bandwidth | 7.7 TB/sData transfer bandwidth between the GPU and local HBM3e memory. |
| FP8 Tensor compute · Dense | 4.5 PFLOPS · DenseTheoretical peak for dense Tensor Core operations. |
| FP8 Tensor compute · Sparse | 9 PFLOPS · SparseTheoretical peak under the applicable structured-sparsity conditions. |
| GPU interconnect | 1.8 TB/s · NVLink 5Aggregate bidirectional GPU-to-GPU bandwidth within compatible HGX systems. |
Before renting, confirm CPU, system memory, storage, networking, deployment region, pricing, and delivery arrangements.
SOC 2 Type II audited
ISO/IEC 27001 certified
GDPR compliant
Compare rental options, explore GPU specifications, then discuss configuration and delivery with our team.
Check the console or contact sales for prices and available instance configurations.
Check the console or contact sales for prices and available instance configurations.
Distinguish GPU specifications, system platforms and the actual delivered configuration.
No. HGX B200 is an eight-GPU Blackwell platform for x86 servers. GB200 NVL72 is a rack-scale system with 72 Blackwell GPUs and 36 Grace CPUs. Their CPU architecture, memory specifications and delivery formats differ.
The HGX B200 column in the NVIDIA datasheet used here lists 7.7 TB/s per GPU; another NVIDIA reference architecture lists up to 8 TB/s. This page keeps the datasheet value. Confirm the exact hardware specification when selecting an instance.
No. The FP8 Dense and Sparse figures are theoretical Tensor Core peaks; Sparse requires the applicable sparsity conditions. FP4 and FP8 also need compatible model formats and kernels. Measure accuracy, latency and throughput with your model and runtime.
Provide B200 quantity, preferred region, start date, duration and single-node or multi-node requirements on the GPU requirements page. Confirm the specific configuration, capacity, access method and quote before ordering.
View configurations, or submit your GPU quantity, region and rental term for a quote.