Memory for large models
GPU memory holds model weights, KV cache and runtime state. Assessing model capacity also requires accounting for precision, context length and concurrency.
Blackwell Ultra architecture with 288 GB of HBM3e per GPU. For large language model inference, fine-tuning and post-training, and distributed multi-GPU training.
Check the console or contact sales for prices and available instance configurations.
B300 combines HBM3e memory, second-generation Transformer Engine and fifth-generation NVLink for AI workloads with substantial compute, memory and communication requirements.
GPU memory holds model weights, KV cache and runtime state. Assessing model capacity also requires accounting for precision, context length and concurrency.
Second-generation Transformer Engine supports low-precision Tensor Core operations. Check model accuracy requirements, frameworks, quantization methods and kernel support before using FP4 or FP8.
Fifth-generation NVLink connects GPUs within compatible HGX systems for data exchange in model-parallel workloads. Cross-node training also depends on separate network configuration and topology.
Blackwell Ultra includes dedicated optimizations for attention computation, targeting reasoning models and long-context workloads. Actual gains depend on the attention implementation, numerical precision and runtime configuration.
From model serving to distributed training, select resources for the workload's memory, compute precision and GPU communication needs.
Plan long-context serving around the 288 GB HBM3e budget for weights and KV cache. Check precision, context length, batch size and concurrent requests together, then measure memory and throughput with the intended reasoning workload.
B300 memory supports weights and training state for compatible fine-tuning or post-training jobs. Include activations, gradients and optimizer state, and plan checkpoint storage; validate the training method and precision against the actual memory budget.
An HGX B300 platform connects eight GPUs with NVLink for model-parallel training. The 1.8 TB/s per GPU is aggregate bidirectional bandwidth within HGX; multi-node training also needs a confirmed network topology and communication benchmark.
Blackwell Ultra optimizes attention computation for long sequences, relevant to image and video models. Size memory by resolution, frame count, sequence length and precision, and verify that the model framework and attention kernels use the supported path.
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 | Blackwell UltraBlackwell Ultra GPU architecture. |
| GPU memory | 288 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.
B300 uses Blackwell Ultra; B200 uses Blackwell. For the configurations presented here, memory is 288 GB and 180 GB per GPU respectively. Memory capacity does not translate directly into training or inference speed; selection also requires precision, software support and workload testing.
No. B300 is a GPU model; HGX B300 is a platform with eight B300 GPUs. The number of GPUs in your rental instance depends on the configuration you select and should be confirmed before renting.
No. HGX B300 is an eight-GPU platform for x86 servers. GB300 NVL72 is a rack-scale system combining Grace CPUs with 72 Blackwell Ultra GPUs. CPU architecture, GPU scale and delivery format differ; their instance specifications are not interchangeable.
No. This figure is the aggregate bidirectional NVLink bandwidth per GPU within compatible HGX systems. Cross-node communication depends on server network adapters, the switching fabric and cluster topology, which must be confirmed separately.
On the GPU requirements page, provide B300 quantity, preferred deployment region, start date, duration and node or network requirements. Then confirm the specific instance configuration, capacity and quote.
View configurations, or submit your GPU quantity, region and rental term for a quote.