Dual DGX Spark Clock Sweep: Finding Efficiency at 1800 MHz

X postWeiResearch

Wei benchmarked a dual DGX Spark setup across clock speeds from 2200 MHz down to 1400 MHz to identify an optimal production profile. Setting the clock to 1800 MHz reduced average power consumption by approximately 30% and peak temperatures by 5–7 °C while sacrificing only minimal performance.

Key points

  1. Clock sweep evaluated frequencies from 2200, 2000, 1800, 1600, down to 1400 MHz.
  2. Compared to 2200 MHz, the 1800 MHz profile retains around 93% single-concurrency decode, 97.5% aggregate decode at four concurrencies, and 93% cold 32K prefill performance.
  3. The decode performance drops by only about 5%, offset by roughly 30% lower average power draw and 5–7 °C lower peak temperatures.
  4. Wei confirmed running the setup actively, noting that the perceived speed remains identical while generating considerably less heat.

Why it matters

In continuous production inference workloads, electricity consumption and thermal load are major drivers of operating costs and system wear. Strategic underclocking illustrates how marginal compute reductions can unlock substantial efficiency gains without degrading real-world user experience.

Analysis

The benchmark highlights diminishing returns at the top end of the hardware clock curve. Pushing hardware to 2200 MHz yields minor latency improvements at the expense of disproportionate power consumption and heat. For inference operations, tuning frequencies through empirical clock sweeps provides a compelling balance between throughput and energy efficiency.

Original post

Wei

@wei_wang · 24 August 2026

把双 DGX Spark 从 2200、2000、1800、1600 一路测到 1400 MHz，终于找到了更适合 Production 的配置。 最终选择 1800 MHz。 在完整 Clock Sweep 里，1800 MHz 相比 2200 MHz 保留了约 93% 的单并发 Decode、97.5% 的四并发 Aggregate Decode，以及 93% 的冷 32K Prefill 性能。 换来的收益很实际。 t.co/3tc2yEfD5r

118 likes33 replies163 bookmarks7.8k views

View on X
More posts in the thread (1)
  1. @Pakero8x yep, running this now, speed feeling the same but with much less heat
Selected replies (4)
  • @shuizhuyu @wei_wang 这种压频测试太实用了，生产环境本来就不追求极限性能，降一点频率换三成功耗和好几度降温，对于长期跑推理的服务来说太划算了。
  • @Pakero8x @wei_wang Pretty confident in this decision after seeing the results at lower frequencies.
  • @QT9277 @wei_wang alpha：能耗比才是真 alpha。
  • @alkaidy2025 @wei_wang 降低约 30% 平均功耗来换5%的decode损失,取舍清晰

Mentioned links

Summary generated by AI (Gemini 3.8 Flash, 27 September 2026). It may contain errors – the original source is authoritative.

Related entries

Similar in content, found by the AI search.

  • X post:Mike Gannotti

    Qwen3.8-Flash-Next on a Single NVIDIA DGX Spark Outscores Dual-Spark Setup in Benchmark

    A benchmark report from SMF Clearinghouse shows MiaAI's Qwen3.8-Flash-Next (NVFP4) scoring 87.3% on smf-bench Official A running on a single NVIDIA DGX Spark (GB10). The single-node serve outperformed the previous dual-Spark DeepSeek V4 Vision-Exp setup in overall accuracy, achieving a perfect score in coding.

    17bookmarks2.3kviews

    AI & LLMs· News

  • X post:Mia

    Performance Update for Qwen3.8 Flash Next on DGX Spark

    Mia (@MiaAI_lab) announced a major performance update for Qwen3.8 Flash Next running on a single DGX Spark, achieving up to 108 tok/s across concurrent streams and extending the KV cache to 1M.

    520bookmarks75.7kviews

    AI & LLMs· Announcement

  • Article:Nemo, DGX Spark & Local Inference Engineer

    Qwen3.8-Flash-Next Scores 87.3% on a Single DGX Spark

    A local benchmark test running Qwen3.8-Flash-Next NVFP4 on a single NVIDIA DGX Spark (GB10) achieved 87.3% on smf-bench Official A, outperforming a previous dual-Spark DeepSeek V4 setup in quality.

    AI & LLMs· Research

  • X post:Tech2Wild

    Optimizing Dual GPU Setups for Large Language Models

    This post explores a technical setup for running large language models using dual GPU configurations, specifically highlighting the use of hardware like the Spark and dual 4090s. The author provides a detailed guide on managing context windows and model instances to improve performance and efficiency for local AI workflows.

    52bookmarks8.7kviews

    AI & LLMs

  • Video

    Video:Heavy Metal Cloud

    Nvidia DGX Spark Hands-On: Setup, vLLM Serving, and Mac Studio Comparison

    Heavy Metal Cloud reviews the Gigabyte and ASUS Nvidia DGX Spark desktop units powered by the Grace Blackwell architecture. The video provides a step-by-step guide for local LLM serving using Docker, vLLM, and SparkRun, as well as media generation in ComfyUI, comparing real-world performance against a Mac Studio.

    AI & LLMs· Tutorial

  • Article:Elite Test Engineering

    Serving DeepSeek V4 Flash on Dual DGX Spark Nodes

    A detailed technical setup report documents serving DeepSeek-V4-Flash (284B MoE) across two DGX Spark nodes powered by Grace Blackwell chips using Ray and vLLM. It highlights workarounds for unified memory monitoring, NCCL RDMA deadlocks, and CUDA graph issues, as well as a 45% speedup using native speculative decoding.

    AI & LLMs· Tutorial

Back to the knowledge base

Let's talk about your project

Get in touch

Locations

  • Mattersburg
    Johann Nepomuk Bergerstraße 7/2/14
    7210 Mattersburg, Austria
  • Vienna
    Ungargasse 64-66/3/404
    1030 Wien, Austria

Parts of this content were created with the assistance of AI.