Evaluating the published long-context path? See whether eight H200 GPUs can run Muse Spark Open, including the prevalidated 1,048,576-token endpoint recipe and its Hopper-specific evidence limits.
Choosing Muse Spark Open hardware? The Muse Spark Open hardware guide separates published eight-B200, four-GB300, eight-H200 and eight-RTX-PRO paths from smaller capacity-only experiments.
Short answer: NVIDIA H200 provides 141GB of HBM3e per GPU, but it is sold as two different data-center deployments: H200 SXM for HGX systems and H200 NVL as a dual-slot air-cooled PCIe product for qualified MGX systems. The capacity can open model and context options beyond 80GB or 96GB, but neither variant should be treated as a drop-in desktop GPU.
Choose the platform before the accelerator
- H200 SXM: a 141GB module for a compatible HGX H200 or DGX-class platform, with power configurable up to 700W.
- H200 NVL: a 141GB dual-slot air-cooled PCIe product for qualified MGX systems, configurable up to 600W.
- A bare listing is not a complete system: confirm baseboard, chassis, airflow, power, firmware and support before comparing price.
H200 SXM vs H200 NVL specifications
| Specification | H200 SXM | H200 NVL |
|---|---|---|
| GPU memory | 141GB HBM3e | 141GB HBM3e |
| Published memory bandwidth | 4.8TB/s | 4.8TB/s |
| Maximum configurable TDP | Up to 700W | Up to 600W |
| Form factor | SXM | PCIe, dual-slot air-cooled |
| Published interconnect | NVLink 900GB/s; PCIe Gen5 128GB/s | 2- or 4-way NVLink bridge, 900GB/s per GPU; PCIe Gen5 128GB/s |
| Named server path | HGX H200 partner and NVIDIA-Certified systems | MGX H200 NVL partner and NVIDIA-Certified systems |
What can 141GB change for local AI?
A 141GB single-device pool can fit artifacts and runtime allocations that exceed 80GB or 96GB. It can also leave more room for KV cache, longer contexts or concurrency after the weights load. Capacity still does not guarantee fit: calculate the exact artifact, cache, workspace and safety margin for your runtime.
Start with the LLM VRAM calculator. For 70B-class workloads, use the 70B VRAM guide to separate raw-weight arithmetic, packaged artifact size and context memory. A workload that fits on H200 is not automatically faster or more economical there than on another platform.
SXM and NVL are different purchases
The H200 SXM module is part of a server platform; it cannot be installed in a normal PCIe slot. H200 NVL uses PCIe and an air-cooled dual-slot form factor, but NVIDIA positions it in qualified MGX systems—not as a general consumer workstation card. “PCIe” describes the interface, not desktop compatibility.
- Verify the exact product label, part number and 141GB capacity.
- For SXM, identify the compatible baseboard, server and cooling solution.
- For NVL, verify the qualified chassis, directed airflow, slot spacing, power mode and cabling.
- Confirm host firmware, driver and management support for the complete system.
Software, MIG and multi-GPU limits
- CUDA target: NVIDIA’s supported-GPU table identifies H200 as Hopper/GH100 with compute capability 9.0. NVIDIA’s MIG setup documentation recommends CUDA 12 for H100/H200. Confirm every framework and compiled extension.
- MIG: both variants can be partitioned into up to seven instances, but NVIDIA lists different maximum slice labels: 18GB for SXM and 16.5GB for NVL on the product page. An assigned slice does not expose all 141GB.
- NVLink: the bridge or fabric does not create one automatic VRAM pool. Software must explicitly partition or parallelize the model; review the multi-GPU VRAM guide.
- Peak throughput: Tensor Core figures marked with sparsity are not comparable to measured tokens per second without matching precision, workload and settings.
Purchase and deployment checklist
- Require label photos and verify SXM versus H200 NVL PCIe; do not rely on a title containing only “H200.”
- Confirm whether the quote includes a tested complete system, a GPU, an SXM module or a multi-GPU NVL configuration.
- Document chassis qualification, airflow, power delivery, circuit capacity, firmware and remote management.
- Validate the exact driver, CUDA, framework, runtime, quantization kernel and container path on a matching system.
- Ask for current device identity, health and memory-error evidence, plus warranty and return terms.
- Compare the whole deployed cost rather than the accelerator price alone.
Alternatives to compare
| Alternative | Capacity | Main comparison |
|---|---|---|
| H100 PCIe, SXM or H100 NVL | 80GB or 94GB | Lower Hopper capacity with multiple form factors |
| RTX PRO 6000 Blackwell | 96GB | Workstation form factor and display-capable professional path |
| DGX Spark / ASUS Ascent GX10 | 128GB unified | Compact complete Arm/Linux system with shared memory |
| Mac Studio | Up to 512GB unified | Apple-silicon software lane and memory shared with the system |
Official NVIDIA sources
- NVIDIA H200 product page and specifications
- NVIDIA MIG supported-GPU table
- NVIDIA H100/H200 MIG software prerequisites
- NVIDIA HGX H200 reference-architecture components
Does the workload need the H200 tier? Compare NVIDIA H100 with H200 for local AI across 80GB, 94GB and 141GB capacity, published bandwidth, power and complete-system requirements.
\nReviewed October 25, 2026. NVIDIA labels product-page specifications preliminary and subject to change. HighVRAM has not independently benchmarked H200. Availability, qualifications and prices change.
Would 180GB per GPU change the workload? Compare NVIDIA H200 with B200 for local AI, separating the 39GB capacity increase and higher published bandwidth from complete HGX platform and software-generation costs.
192GB AMD comparison: use the MI300X local-AI guide to compare capacity, platform and software requirements without inferring speed from specifications.
\n\nNeed the 192GB AMD alternative? Compare NVIDIA H200 with AMD MI300X across Qwen3.5 fit, ROCm/CUDA, multi-accelerator topology and platform requirements.
For the current Blackwell Ultra capacity step, review the NVIDIA B300 288GB model-fit guide.
Looking beyond H200? Compare NVIDIA B200 and B300 for local AI across usable memory, exact large-model fit, software support and platform requirements.
Considering Blackwell after H200? Review the NVIDIA B200 180GB local-AI guide for exact model fit, CUDA readiness and system requirements.
Need the full NVIDIA generation ladder? Compare H100, H200, B200 and B300 with exact large-model fit and system tradeoffs.
Desktop GPU or 141GB Hopper accelerator? Compare H200 SXM/NVL with RTX 5090 32GB across exact model-fit, cooling, power, MIG and whole-system requirements.
DGX Spark alternatives: See when H200 is—and is not—a practical DGX Spark alternative in the constraint-first comparison hub.
DGX Spark vs H200: Compare H200 directly with DGX Spark 128GB shared memory in the exact 128GB-versus-141GB comparison.