141 GB HBM3e memory
More memory than the 80 GB H100 SXM reference gives additional space for model weights, KV cache and training state. Usable model capacity still depends on precision, context and concurrency.
Hopper architecture with 141 GB of HBM3e and 4.8 TB/s of memory bandwidth per GPU. This page uses H200 SXM reference specifications for model inference, fine-tuning and multi-GPU training.
Check the console or contact sales for prices and available instance configurations.
H200 SXM combines HBM3e memory with Hopper Tensor Cores and fourth-generation NVLink. Evaluate memory capacity, low-precision software support and communication for the target workload.
More memory than the 80 GB H100 SXM reference gives additional space for model weights, KV cache and training state. Usable model capacity still depends on precision, context and concurrency.
The HBM3e interface moves model parameters and working data between memory and compute. Its value to inference or scientific computing depends on the workload's memory access pattern.
Fourth-generation Tensor Cores and Transformer Engine support FP8 and 16-bit operations. Framework, kernel and model accuracy support must be checked before selecting lower precision.
H200 SXM provides 900 GB/s of aggregate bidirectional NVLink bandwidth per GPU in compatible systems. Multi-node jobs also require separately specified network adapters and topology.
From model serving to distributed training, select resources for the workload's memory, compute precision and GPU communication needs.
For serving dominated by weights and working data, assess 141 GB HBM3e and 4.8 TB/s against the 80 GB H100 SXM reference. Select GPU count from the full memory requirement and benchmark the actual access pattern.
Reserve KV cache within the 141 GB budget after weights and runtime buffers. Context length and concurrent sequences compete for that space, so test their combined limits before setting request capacity.
Budget activations, gradients and optimizer state rather than model weights alone. Adapter tuning and full fine-tuning have different state requirements; validate memory use and FP8 software support for the selected method.
Use the H200 SXM reference of 900 GB/s bidirectional NVLink per GPU to assess communication within a compatible node. For training across nodes, separately confirm the cluster fabric and how the parallelism strategy uses it.
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 | HopperHopper GPU architecture. |
| GPU memory | 141 GB HBM3eStores model weights, KV cache and runtime state. |
| Memory bandwidth | 4.8 TB/sData transfer bandwidth between the GPU and local HBM3e memory. |
| FP8 Tensor compute · Dense | 1,979 TFLOPS · DenseTheoretical peak for dense Tensor Core operations. |
| FP8 Tensor compute · Sparse | 3,958 TFLOPS · SparseTheoretical peak under the applicable structured-sparsity conditions. |
| GPU interconnect | 900 GB/s · NVLink 4Aggregate 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.
Both use Hopper. H200 SXM provides 141 GB HBM3e and 4.8 TB/s of memory bandwidth; the H100 SXM reference provides 80 GB HBM3 and 3.35 TB/s. The additional memory can change model placement and batch sizing, but does not imply the same speedup for every application.
No. Both list 141 GB and 4.8 TB/s, but SXM and NVL use different form factors and have different compute and power specifications. NVIDIA lists 3,958 TFLOPS FP8 Sparse for SXM and 3,341 TFLOPS for NVL. This page uses SXM values; confirm the delivered variant and GPU topology.
The datasheet's 3,958 TFLOPS is a theoretical peak with sparsity. The 1,979 TFLOPS Dense figure is derived from NVIDIA's Hopper description of a 2× structured-sparsity uplift. Sparse performance requires the applicable conditions, and FP8 needs compatible software and accuracy validation. Neither figure predicts model throughput.
Specify H200 SXM, GPU quantity, preferred region, start date, duration and node or network requirements on the GPU requirements page. Confirm the exact variant, capacity, configuration, access method and quote before ordering.
View configurations, or submit your GPU quantity, region and rental term for a quote.