NVIDIA DGX Station benchmarks

Reproducible local ML training, operator, and LLM inference results from NVIDIA DGX Station systems with one NVIDIA GB300 selected per station. Multi-node experiments use generic node0 and node1 roles. These are GB300 systems, not DGX Spark/GB10.

The two-station direct-connect test setup.

DGX Station guide

Start with the DGX Station GB300 field guide for measured system specifications, photos, two-node networking, container setup, performance tuning, runtime quirks, benchmarking practice, and safe recovery.

Experiments

GLM-5.3 headline

Full GLM-5.3 results, proposal sweep, and recipes →

Each experiment folder contains:

• A complete README with benchmark conditions, tables, quality results, and caveats
• Embedded publication-ready graphs
• Machine-readable CSV data and experiment-specific quality/audit JSON
• An agent-ready recipes/ directory with pinned setup and reproduction commands

Test system

The display GPU was excluded from inference. Results are direct measurements from the tested systems, not vendor projections.

Shared methodology

The throughput measurements use llm-inference-bench v0.4.29 at commit 0b4185b5b435e948b199c9077a00b084864aa963. Qwen and DeepSeek use its finite-request layer:

• 8,192 input tokens and 1,024 generated tokens
• Temperature 0, EOS ignored for a fixed amount of decode work
• Concurrency 1, 2, 4, 8, 16, 32, 64, and 128
• 5 × concurrency measured requests after concurrency warm-up requests
• Aggregate output throughput = measured output tokens / benchmark wall time

Quality was tested separately with EOS respected, experiment-specific natural or mixed prompts, canonical WikiText-2 perplexity, and automated repetition audits. See each experiment README before comparing numbers; prompt construction, cache precision, and model architecture differ.

DeepSeek-V4.1-Flash prefill does not use llm-inference-bench. Its prefill tables come from that section's own bench_prefill.py client against each engine's native endpoint: unique random-token prompts of exactly 16K, 32K, 64K, and 128K tokens, one generated token, temperature 0, a fixed 1, 4, or 16 requests held in flight, a cache flush before every point, and aggregate prompt tokens divided by the wave's wall time. Those multi-request aggregate values are not interchangeable with the single-request cold-prefill cells of the llm-inference-bench sections. Its decode rows use the finite-request layer above with the same pinned client.

MiniMax H3 uses a separate fixed-seed video-and-audio methodology: one full 50-step warmup followed by three measured 1344×768, 124-frame requests. Its folder also measures native 10- and 15-second samples plus one explicitly unsupported patched 30-second attempt. It reports end-to-end latency, stage timing, peak HBM, media integrity, and manual non-degeneracy review; its results are not comparable to LLM token rates.

Ornith, GLM-5.2, and Hy3 instead use the benchmark's fixed-duration sustained-decode layer: offered concurrency is maintained during a 30-second measurement window, with separately labeled 60-second stability cells where present. Boundary requests may remain in flight when the window closes. Sustained-window throughput is not directly interchangeable with the finite 5 × concurrency results above; each experiment README identifies its layer and retains completion and scheduler-residency fields.

Contents

.
├── qwen3.8-flash-next/
│   ├── README.md
│   ├── charts/
│   ├── data/
│   └── recipes/
├── qwen3.8-27b/
│   ├── README.md
│   ├── charts/
│   ├── data/
│   └── recipes/
├── nanogpt-training/
│   ├── README.md
│   ├── data/
│   └── recipes/
├── mlperf-retinanet-training/
│   ├── README.md
│   ├── data/
│   └── recipes/
├── gdn2-linear-attention/
│   ├── README.md
│   ├── data/
│   └── recipes/
├── gdn2-mamba3-te-comparison/
│   ├── README.md
│   ├── charts/
│   ├── data/
│   └── recipes/
├── transformer-engine-training/
│   ├── README.md
│   ├── data/
│   └── recipes/
├── mamba3-training/
│   ├── README.md
│   ├── data/
│   └── recipes/
├── dgx-station-guide/
│   ├── README.md
│   └── photos/
├── deepseek-v4-flash-0731/
│   ├── README.md
│   ├── charts/
│   ├── data/
│   └── recipes/
├── deepseek-v4.1-flash/
│   ├── README.md
│   ├── charts/
│   ├── data/
│   ├── notes/
│   └── recipes/
├── ornith-1.5-397b/
│   ├── README.md
│   ├── charts/
│   ├── data/
│   └── recipes/
├── glm-5.2/
│   ├── README.md
│   ├── charts/
│   ├── data/
│   └── recipes/
├── glm-5.3/
│   ├── README.md
│   ├── charts/
│   ├── data/
│   └── recipes/
├── glm-5.3-flash/
│   ├── README.md
│   ├── charts/
│   ├── data/
│   └── recipes/
├── hy3/
│   ├── README.md
│   ├── charts/
│   ├── data/
│   └── recipes/
├── minimax-h3-video/
│   ├── README.md
│   ├── charts/
│   ├── data/
│   └── recipes/
├── minimax-m3/
│   ├── README.md
│   ├── charts/
│   ├── data/
│   └── recipes/
└── gb300-networking/
    ├── README.md
    ├── charts/
    ├── data/
    └── recipes/

Measured August–September 2026.