Cosmos3-Super serving study: B200 replica topology and H200/B200 parallelism

[stewtong]

I published a measurement record and source-visible serving implementation for nvidia/Cosmos3-Super text-to-video generation on one eight-GPU H200 node and one eight-GPU B200 node using vLLM-Omni:

https://github.com/stewtong/cosmos3-super-serving

Measurement context: August 2026, BF16, 1280 x 720, 189 frames at 24 fps, 35 denoising steps, vLLM 0.25.0, and model snapshot e0262be9d8f7586bc24c069a2aed2b665bdff266. The B200 topology run used driver 580.173.02; the H200 and other B200 observations used driver 580.126.09. The repository contains the full H200 and B200 environment records.

NVIDIA's Cosmos3-Super Generator text-to-video grid reports latency across one, four, and eight GPUs. This record adds whole-node topology and serving observations:

• Across four primary B200 topology cells with 24 attempts per cell, concurrency one, and guardrails off, eight one-GPU services produced 43.2% more technically valid video-seconds per node-hour than one eight-GPU service, at 5.54 times the median client-wall latency. Technical validity covers decoding, dimensions, frame count, frame rate, duration, and blank/frozen detection; semantic quality and human acceptance were not measured.
• The H200 and B200 serving summaries record same-host comparisons between TP-8 and the CFG-2 x Ulysses-4 x HSDP-8 layout recommended for H200. At 35 steps with guardrails off, H200 mean server-generation time was 120.57 seconds versus 131.26 seconds for TP-8, based on three runs per layout. B200 mean client-wall latency was 69.0 seconds versus 77.3 seconds for TP-8, also based on three runs per layout. These are recorded observations because their raw per-request inputs are not included.
• The H200 serving summary reports that concurrency one to eight increased throughput from 28.21 to 30.96 clips per hour while p50 latency rose from 127.6 to 923.3 seconds and reported peak GPU memory moved from 38,562 to 38,564 MB. This guardrails-on, random-prompt pattern is consistent with queueing rather than batching. Its raw per-request inputs are not included.

The B200 topology record contains 96 primary-cell attempts, 48 concurrency-two confirmation attempts, and three drift-check attempts. It includes exact measurement windows, output hashes, technical-validity results, and derivation code. The source MP4s are excluded, so the recorded hash matches and validity outcomes cannot be independently rerun from this repository.

The four primary topology cells held the workload, prompt hashes, seed cycle, guardrails-off posture, and concurrency one constant, and applied the same technical-validity gate to every attempt. The concurrency-two confirmations are labeled separately. NVIDIA's public grid does not disclose the exact image, model revision, or driver used for its values, and I did not reproduce that environment. Cross-project latency comparisons are therefore contextual.

Hardware context: the measurements ran on Nebius H200 and B200 nodes. The linked repository is public under default copyright with no software license.

Would a clearly labeled community-reference link under the Cosmos3-Super Generator notes be useful? I can submit a focused documentation PR if so.