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.
Short answer: NVIDIA B200 provides 180GB of HBM3E on one Blackwell SXM accelerator. That is enough by file size for every listed Qwen3.5-122B artifact through Q8, but not the 250.173GB BF16 checkpoint. For Qwen3.5-397B, even the 235.708GB GPTQ 4-bit files exceed one B200, so the decision immediately becomes multi-GPU placement, a larger-capacity accelerator, or offload—not a single-card fit.
B200 is not a desktop graphics card. NVIDIA’s enterprise reference describes eight 180GB SXM GPUs on an HGX B200 baseboard. A complete qualified server, rack power, cooling, firmware, CUDA stack, networking and support are part of the product decision. Treat bare-module marketplace listings with caution.
NVIDIA B200 specifications that matter for local AI
| Architecture | NVIDIA Blackwell / GB100 |
|---|---|
| Memory per GPU | 180GB HBM3E |
| Published memory bandwidth | Up to 8TB/s per GPU |
| Compute capability | 10.0 |
| Interconnect | Fifth-generation NVLink and NVSwitch |
| GPU-to-GPU bandwidth | Up to 1.8TB/s per GPU |
| Eight-GPU aggregate memory | 1.44TB across separate device memories |
| Aggregate NVLink bandwidth | 14.4TB/s |
| Maximum configurable power | Up to 1,000W per GPU |
| Multi-instance GPU | Up to seven MIG instances |
What fits on one B200: Qwen3.5-122B-A10B
| Artifact | Files | Margin in 180GB | Capacity result |
|---|---|---|---|
| GPTQ 4-bit | 78.864GB | 101.136GB | Fits by files |
| Q4 plus projector | 77.446GB | 102.554GB | Fits by files |
| Official FP8 | 127.163GB | 52.837GB | Fits by files |
| Q8 plus projector | 130.781GB | 49.219GB | Fits by files |
| Official BF16 | 250.173GB | Does not fit | Arithmetic minimum: 2 × 180GB |
File margin is not free runtime memory. CUDA context, cache, hybrid-attention state, temporary buffers, multimodal processing, batch and concurrency all consume capacity. Measure peak device use with the intended runtime and context before treating any row as deployable.
What fits on B200: Qwen3.5-397B-A17B
| Artifact | Files | One B200? | Arithmetic minimum |
|---|---|---|---|
| GPTQ 4-bit | 235.708GB | No | 2 × 180GB = 360GB |
| Q4 plus projector | 245.012GB | No | 2 × 180GB = 360GB |
| Official FP8 | 406.152GB | No | 3 × 180GB = 540GB |
| Q8 plus projector | 422.426GB | No | 3 × 180GB = 540GB |
| Official BF16 | 806.796GB | No | 5 × 180GB = 900GB |
Those totals are division, not deployment recommendations. NVLink and NVSwitch improve communication, but they do not present several B200 memories as one transparent 360GB, 540GB or 900GB device. The framework must support the intended tensor, pipeline, expert or other parallel placement, with headroom on every GPU.
B200 vs H200 and B300
| GPU | Architecture | Memory | Published bandwidth | Planning distinction |
|---|---|---|---|---|
| H200 SXM | Hopper | 141GB HBM3E | 4.8TB/s | Smaller mature Hopper tier |
| B200 SXM | Blackwell | 180GB HBM3E | Up to 8TB/s | More single-device room than H200 |
| B300 SXM | Blackwell Ultra | 288GB product tier; 270GB individual-HGX datasheet value | Up to 8TB/s tier; 7.7TB/s individual-HGX datasheet | Crosses selected 235–250GB single-device boundaries |
B200’s 39GB advantage over H200 can matter at exact boundaries, but B300 changes a different class of decision: conservatively planned at 270GB, it can hold the listed Qwen3.5-397B 4-bit files or Qwen3.5-122B BF16 files on one device. Compare the exact artifact, runtime headroom, software readiness, complete-system cost and measured workload.
Blackwell software readiness is part of fit
NVIDIA identifies B200 as compute capability 10.0. A host driver recognizing the GPU does not prove that the full application path is ready. Confirm the exact driver, CUDA toolkit, framework or serving runtime, container image, quantization kernels, communication libraries and compiled extensions. For older or custom software, run the intended model and measure correctness, peak memory and performance before procurement.
MIG changes the visible memory
NVIDIA lists up to seven MIG instances for B200. A MIG workload sees its assigned slice, not the physical GPU’s complete 180GB. Record whether MIG is enabled, the profile, memory exposed to the process and any other tenants before comparing results or claiming a model fits.
Complete-system and used-listing checklist
- Identify a complete qualified B200 server or the exact supported baseboard—not just a loose SXM module.
- Verify GPU count, 180GB exposed per accelerator, NVLink/NVSwitch topology and any MIG partitioning.
- Confirm rack power, cooling, host CPUs, system memory, storage, networking and service requirements.
- Pin the driver, CUDA, container, framework, kernels and model revision.
- Measure peak per-device memory, context, batch, concurrency, time to first token, throughput, throttling and wall power.
- Compare complete acquisition, support and operating cost with H200, B300 and other qualified alternatives.
Plan the next step
- Compare H200 and B200
- Compare B200 and B300
- Review the B300 capacity guide
- Build a reproducible benchmark record
- See Qwen3.5-122B capacity tiers
- See Qwen3.5-397B capacity tiers
- Compare all large-model hardware paths
- Browse the hardware guide
- Follow the high-VRAM watchlist
Primary sources
- NVIDIA HGX H100/H200/B200 enterprise-reference components
- NVIDIA DGX B200 specifications
- NVIDIA MIG supported-GPU table
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’s current documentation before purchase.
Need a higher-capacity cross-vendor option? Compare B200 180GB with AMD MI325X 256GB for exact Qwen3.5 boundaries, CUDA versus ROCm readiness and qualified-system constraints.
Upgrading from Hopper? Compare H100 and B200 for local AI using exact capacity, software-generation and complete-system tradeoffs.
Choosing among four NVIDIA datacenter tiers? Use the H100/H200/B200/B300 decision hub for one capacity and platform comparison.