Djangodevreng/dgx-spark-benchmarks

DGX Spark LLM Arena benchmarks

Reproducible LLM inference benchmarks on an NVIDIA DGX Spark (GB10, 128 GB unified memory). The suite defines eleven tests: six closed-loop (llama-benchy) and five open-loop (vllm bench serve). Results cover all eleven: the ten throughput tests under results, and the rate sweep under rateSweep. Raw results remain inspectable, but only complete runs without a failed sanity check count toward rankings and aggregate throughput. Open-loop tests must additionally complete at least 99% of requests. reliabilityReason explains every exclusion. Rate-sweep capacity reports sustained throughput in requests per second under a p95 TTFT ceiling of 2, 5 and 10 seconds (null on a threshold means the model never met it). abortedAtRps is non-null when the server died partway through the sweep: the steps above that rate were never measured, so the capacity is a lower bound and not a measured ceiling. queueKneeRps is the lowest rate after which peak concurrency more than doubles while throughput rises by less than a quarter: the point where added load buys queueing instead of work. Legacy scores use only measured throughput and throughput per GB VRAM. The separate suitabilityScores are indicative quality-led scores with a strict benchmark reference; see suitabilityScoringPolicy. Quality metrics are external vendor model-card figures, not local precision validation. Every metric retains its harness because MMLU and MMLU-Pro, GPQA variants and coding evals are not interchangeable.

• Hardware: NVIDIA DGX Spark (GB10, 128 GB unified memory)
• Creator: Django de Vreng (https://djangodevreng.nl)
• Visualised: https://djangodevreng.nl/arena/
• Raw runs: https://github.com/djangodevreng/dgx-spark-benchmarks
• Last updated: 2026-08-17

Configs

• models: one row per model (38 rows). Metadata, quality scores and the leaderboard composite scores per use-case preset.
• results: long format, one row per measured (model, benchmark) pair. Covers the 10 throughput tests.
• capacity: long format, one row per (model, latency threshold). The rate sweep, reported separately because it measures a curve rather than a single point.

The suite has 11 benchmarks: 6 closed-loop (llama-benchy) and 5 open-loop (vllm bench serve). All 11 are represented — 10 in results, the rate sweep in capacity.

Capacity columns

capacity reports the highest sustained request rate that stayed under a p95 TTFT ceiling, at three thresholds: 2000, 5000 and 10000 ms. Empty fields mean the model never met that threshold at any step of the sweep — that is a result, not missing data. configured_rps is the rate that was offered, achieved_rps what the server actually sustained; the gap between them shows where a model starts falling behind.

Read sweep_aborted_at_rps before using a capacity figure. When it is set, the server died at that rate and the steps above it were never measured, so the capacity for that model is a lower bound rather than a measured ceiling. It repeats on all three rows of a model because it qualifies each of them.

queue_knee_rps is the lowest rate after which peak concurrency more than doubles while throughput rises by less than a quarter: the point where added load buys queueing instead of work. Empty means no such knee was detected in the measured range.

Context columns

context_k is the model's native context window as published by the vendor. measured_context_k is the max_model_len the server actually ran with, which is what these throughput figures rest on. They are usually not the same: the suite standardises on 128K so models stay comparable, and a few runs sit lower still because the model or the memory did not allow more. Read throughput against measured_context_k, not against the spec.

Quality columns

Quality is reported on three dimensions, each with the benchmark the number came from: knowledge / knowledge_bench, science / science_bench, coding / coding_bench. The underlying benchmark differs per model — knowledge is MMLU-Pro for most models but plain MMLU for some, and coding ranges from LiveCodeBench to HumanEval+ — so always read the value together with its _bench column. Figures come from vendor model cards, not from own testing. quality_avg is the mean across the three dimensions.

from datasets import load_dataset

models = load_dataset("Djangodevreng/dgx-spark-benchmarks", "models")
results = load_dataset("Djangodevreng/dgx-spark-benchmarks", "results")

Method

Closed-loop benchmarks run three times per measurement point and are reported as the mean and standard deviation; open-loop benchmarks run once with a fixed seed (42). Variance is reported per measurement rather than hidden behind one global bound. No latency gate: slow models stay visible. Full methodology and the raw stdout per run live in this repo.

License

CC-BY-4.0. Free to use, including commercially, with attribution: Django de Vreng, https://djangodevreng.nl.

Citation

@misc{devreng-dgx-spark-benchmarks-2026,
  author = {Django de Vreng},
  title  = {DGX Spark LLM Arena benchmarks},
  year   = {2026},
  url    = {https://huggingface.co/datasets/Djangodevreng/dgx-spark-benchmarks}
}