Reproducing the official self-managed path? See whether four GB300 GPUs can run Muse Spark Open, including the published SGLang TP4/EP4 configuration and GB300/B300 naming boundary.
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.
Following the published 768GB path? See whether eight RTX PRO 6000 Blackwell GPUs can run Muse Spark Open, including the prevalidated 131,072-context recipe, 96GB/98GB source discrepancy and validation limits.
Evaluating the 720GB boundary? See whether four B200 GPUs can run Muse Spark Open and why passing the 700GB arithmetic does not reproduce the published eight-B200 node path.
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.
New 700GB-class workload: the Muse Spark Open VRAM guide separates its exact 533.011GB NVFP4 files from Meta’s official 700GB minimum and compares 640GB, 705GB, 720GB and 768GB pools.
Current 640GB boundary: the 8×H100 80GB Kimi K2.5 guide compares that tight fit with H200, B200, 96GB and MI325X paths.
Current Kimi workload: use the Kimi K2.5 hardware guide to compare 8×80GB, 7×94GB, 5–6×H200, 4×B200, 8×96GB and B300-class capacity arithmetic.
Current distributed workload: use the Kimi-K2.5 VRAM guide to compare 8×80GB, 4×180GB and 8×96GB NVIDIA capacity paths.
Planning MiniMax-M2.5? Use the 230.134GB checkpoint guide before translating NVIDIA aggregate capacity into a placement plan.
Planning DeepSeek-V3.2? Use the 689.483GB checkpoint hardware guide to compare H200, B200/B300, 8×96GB and AMD alternatives without treating aggregate capacity as a deployment guarantee.
Short answer: NVIDIA’s H100, H200, B200 and B300 datacenter tiers move from 80GB Hopper through 141GB Hopper, 180GB Blackwell and a 288GB-class Blackwell Ultra product whose individual HGX datasheet value is 270GB. The useful choice is not “newest wins.” Choose the smallest qualified platform that fits the complete workload with measured headroom, supported software and acceptable system cost, power and cooling.
This page compares SXM-class HGX tiers. H100 and H200 also have NVL PCIe variants, and H100 has an 80GB PCIe product. Those form factors are covered in the detailed product guides. Aggregate node memory below remains distributed across separate devices; NVLink/NVSwitch does not create one transparent memory pool.
H100 vs H200 vs B200 vs B300 specifications
| GPU tier | Architecture | Memory per GPU | Published bandwidth | GPU-to-GPU bandwidth | Maximum configurable power |
|---|---|---|---|---|---|
| H100 SXM | Hopper / capability 9.0 | 80GB HBM3 | 3.35TB/s | 900GB/s | Up to 700W |
| H200 SXM | Hopper / capability 9.0 | 141GB HBM3E | 4.8TB/s | 900GB/s | Up to 700W |
| B200 SXM | Blackwell / capability 10.0 | 180GB HBM3E | Up to 8TB/s | Up to 1.8TB/s | Up to 1,000W |
| B300 SXM | Blackwell Ultra | 288GB product tier; 270GB individual-HGX datasheet value | Up to 8TB/s tier; 7.7TB/s individual-HGX datasheet | Up to 1.8TB/s | Up to 1,100W |
Qwen3.5-122B-A10B: single-device file fit
| Artifact | Files | H100 80GB | H200 141GB | B200 180GB | B300 270GB planning |
|---|---|---|---|---|---|
| GPTQ 4-bit | 78.864GB | 1.136GB margin; unsafe boundary | 62.136GB margin | 101.136GB margin | 191.136GB margin |
| Q4 plus projector | 77.446GB | 2.554GB margin; unsafe boundary | 63.554GB margin | 102.554GB margin | 192.554GB margin |
| Official FP8 | 127.163GB | No | 13.837GB margin; tight | 52.837GB margin | 142.837GB margin |
| Q8 plus projector | 130.781GB | No | 10.219GB margin; tight | 49.219GB margin | 139.219GB margin |
| Official BF16 | 250.173GB | No | No | No | 19.827GB margin; tight |
“Margin” subtracts artifact files from nominal device capacity. It does not reserve CUDA context, cache, hybrid-attention state, temporary buffers, projector/vision processing, batch or concurrency. H100’s 1–3GB remainder and H200’s 10–14GB FP8/Q8 remainder are particularly unsafe to treat as deployment proof.
Qwen3.5-397B-A17B: arithmetic device counts
| Artifact | Files | H100 80GB | H200 141GB | B200 180GB | B300 270GB planning |
|---|---|---|---|---|---|
| GPTQ 4-bit | 235.708GB | 3 leaves 4.292GB aggregate; plan 4 | 2 | 2 | 1; 34.292GB file margin |
| Q4 plus projector | 245.012GB | 4 | 2 | 2 | 1; 24.988GB file margin |
| Official FP8 | 406.152GB | 6 | 3 leaves 16.848GB aggregate; tight | 3 | 2 |
| Q8 plus projector | 422.426GB | 6 | 3 totals 423GB and leaves only 0.574GB; plan 4 | 3 | 2 |
| Official BF16 | 806.796GB | 11 | 6 | 5 | 3 totals 810GB and leaves only 3.204GB; plan 4 |
These counts are transparent division, not validated deployments. A runtime must explicitly place or shard the model and its state across GPUs, and every device needs headroom. The H100 GPTQ, H200 Q8 and three-B300 BF16 rows are file-size traps even though aggregate arithmetic clears the artifact total.
Choose the tier by the boundary it removes
- H100 80GB: relevant when the measured workload stays well below 80GB or an existing Hopper platform is already available. Do not plan around a 1–3GB file remainder.
- H200 141GB: keeps Hopper/capability 9.0 while moving 122B FP8 and Q8 into tight one-device file fits.
- B200 180GB: adds safer single-device room for 122B FP8/Q8 and reduces several 397B arithmetic device counts, with a Blackwell software-validation step.
- B300 270GB planning: crosses the listed 122B BF16 and 397B 4-bit single-device file boundaries, but both retain finite runtime headroom.
Software continuity versus capacity
H100 and H200 are Hopper with compute capability 9.0. B200 is Blackwell with capability 10.0; B300 is Blackwell Ultra. A working Hopper container does not prove that its framework build, quantization kernels, communication libraries or compiled extensions support Blackwell correctly. Validate the exact driver, CUDA toolkit, container, framework, serving runtime and model-specific kernels on the candidate system.
Power, cooling and acquisition are platform decisions
The per-GPU maximum moves from up to 700W on H100/H200 SXM to up to 1,000W on B200 and 1,100W on B300. Compare complete qualified servers: GPU count, baseboard, cooling, rack power, host CPUs, system memory, storage, networking, firmware, service, delivery and replacement terms. A loose SXM module is not a workstation upgrade path.
Reproducible comparison checklist
- Identify the exact system SKU, GPU count and memory exposed per accelerator.
- Record MIG state, NVLink/NVSwitch topology, placement and any offload.
- Pin driver, CUDA, container, framework, kernels and model revision.
- Hold artifact, precision, context, batch/concurrency and quality target constant.
- Measure peak per-device memory, time to first token, prompt and generation throughput, throttling and wall power.
- Compare complete acquisition, support and operating cost—not specification ratios alone.
Detailed guides and next actions
- H100 PCIe, SXM and NVL guide
- H200 SXM and NVL guide
- B200 180GB guide
- B300 capacity guide
- H100 vs B200 detailed comparison
- B200 vs B300 detailed comparison
- Build a reproducible benchmark record
- Compare large-model hardware paths
- Browse the hardware guide
- Follow the high-VRAM watchlist
Primary sources
- NVIDIA H100 Tensor Core GPU specifications
- NVIDIA HGX H100/H200/B200 enterprise-reference components
- NVIDIA HGX H200/B200/B300 enterprise-reference components
- NVIDIA Blackwell Ultra datasheet
Capacity calculations use decimal GB and compare artifact files with nominal device capacity. They are not runtime-memory guarantees. Product specifications and software support can change; confirm the exact system before purchase.
Cross-vendor boundary: compare NVIDIA B200 180GB with AMD MI325X 256GB when CUDA/Blackwell readiness and a larger ROCm/gfx942 memory tier are both viable.
Need the 48GB entry point below the Hopper ladder? Compare L40S with H100 80GB/94GB before moving through H200, B200 and B300.
Desktop-to-datacenter boundary: use the RTX 5090 vs H200 comparison for exact 32GB-versus-141GB Qwen3.5 margins and SXM/NVL safeguards.
DGX Spark vs H100: Add DGX Spark to the NVIDIA data-center tier decision in the exact capacity and platform comparison.
DGX Spark vs H200: Add DGX Spark to the H100/H200 tier choice in the exact 128GB-versus-141GB comparison.
The A100 tier now has a direct compact-system comparison: DGX Spark vs A100 40GB/80GB, with PCIe/SXM, cooling and MIG safeguards.
Current workload path: use the GLM-4.7-Flash hardware guide to map its 62.444GB official BF16 checkpoint onto H100, H200, B200 and B300 capacity tiers.
Current distributed workload: the DeepSeek-V3.2 VRAM guide applies the exact 689.483GB official checkpoint to 640GB, 768GB and larger NVIDIA capacity plans.