Lucebox (AMD Radeon AI PRO R9700 + Strix Halo) Beats NVIDIA DGX Spark by 3.63x on DeepSeek V4 Flash Decode Speed | lucebox

By Davide Ciffa<br>Lucebox (AMD Radeon AI PRO R9700 + Strix Halo) Beats NVIDIA DGX Spark by 3.63x on DeepSeek V4 Flash Decode Speed<br>51.1 tok/s (Lucebox) vs 14.09 tok/s (single DGX Spark)<br>DeepSeek V4 Flash decode comparison<br>Measured decode throughput for one request at a time on the full 284B model SystemTested configurationDecode throughput LuceboxROCmFPX compressed model, DSpark draft model, four experts per token51.1 tok/s median 1&times; NVIDIA DGX SparkQ2 compressed model, average of tests at four context lengths14.09 tok/s Measured ratio51.1 &divide; 14.093.63&times; Using the unrounded DGX Spark mean of 14.0875 tok/s gives the 3.63&times; headline ratio. Each platform runs a configuration tuned for its own hardware, so this is a whole-system comparison rather than a GPU-only A/B test. Decode throughput is the rate at which each system generates new tokens.<br>DeepSeek V4 Flash has 284 billion parameters. The 102.3 GB ROCmFPX build is too large for the R9700&rsquo;s 32 GB of memory, so Lucebox gives the two AMD GPUs different jobs. The R9700 handles the dense path and frequently selected experts; Strix Halo holds the remaining experts in its 128 GB memory and runs them in parallel.<br>The headline Lucebox result is the median of three measured requests after two warmups, using an approximately 2k-token prompt and 128 generated tokens. An 11.3 GB DSpark draft model proposes several tokens at once; the full DeepSeek V4 Flash model verifies them before they are accepted. We also ran a separate 53-token prompt to measure the upper end of the implementation.<br>For the single-device baseline and the ROCmFPX compression details, see our companion report on DeepSeek V4 Flash on AMD Ryzen AI MAX+ 395.<br>Complete system price<br>AMD-Powered Lucebox is priced at $6,499 as a complete local inference system, including the custom chassis, 2 TB of storage, power delivery, hardware integration, validation, and warranty. NVIDIA lists one DGX Spark at $4,699, or $9,398 for two.<br>Complete systemCurrent U.S. priceWhat is included 1&times; DGX Spark$4,699Complete system with 4 TB storage AMD-Powered Lucebox$6,499Custom chassis, 2 TB storage, power delivery, integration, validation, and warranty 2&times; DGX Spark$9,398Two complete systems At those list prices, Lucebox costs 38% more than one DGX Spark and delivers about 2.6&times; the decode throughput per dollar in the headline configurations. It costs 31% less than two DGX Sparks.<br>Prices are the U.S. list prices available at publication, before tax and shipping. Sources: Lucebox system configuration and NVIDIA Marketplace. Configurations and prices can change.<br>One DGX Spark, measured directly<br>We ran a Q2 compressed version of the full DeepSeek V4 Flash model on one DGX Spark. We then measured Lucebox at the same 2k, 4k, 8k, and 16k context lengths.<br>Matched single-request decode throughput on the full 284B model ContextLucebox1&times; DGX SparkSpeedup 2k51.0 tok/s14.18 tok/s3.60&times; 4k49.6 tok/s14.24 tok/s3.48&times; 8k47.5 tok/s14.04 tok/s3.38&times; 16k42.9 tok/s13.89 tok/s3.09&times; Mean 47.75 tok/s 14.09 tok/s 3.39&times; The DGX Spark mean is 14.0875 tok/s, which we report as 14.09 tok/s. Lucebox averages 47.75 tok/s across the same four context lengths, a 3.39&times; speedup. The 3.63&times; headline uses the separate 51.1 tok/s Lucebox serving result shown above.<br>DGX Spark ran the Q2 model; Lucebox ran the ROCmFPX model with four experts per token and DSpark speculative decoding. Each system uses the configuration designed for its hardware.<br>For wider context, LocalMaxxing includes two-DGX-Spark results both with the standard target and with a DSpark-specific package. We show both below rather than selecting the more favorable reference.<br>DeepSeek V4 Flash decode, one request at a time · tok/s<br>1&times; DGX Sparkaverage of four tests

14.09

2&times; DGX Sparkwithout a helper model

45.70

AMD-Powered Luceboxserving median

51.10

AMD-Powered Luceboxshort prompt, fastest run

55.00

2&times; DGX Sparkwith DSpark

65.09

The Lucebox values are a three-request serving median and the fastest of three separate runs with a 53-token prompt. The single DGX Spark figure is our mean across 2k, 4k, 8k, and 16k context tests. Public sources: two DGX Sparks without a draft model and two DGX Sparks with DSpark. Prompts, context lengths, software, compression, and routing settings differ, so the public runs provide context rather than a controlled A/B comparison.

The public two-Spark results bracket Lucebox: the standard target is effectively tied with our longer decode run, while the DSpark-specific configuration is faster. They are useful reference points, but not substitutes for a matched benchmark.<br>How the work is split<br>Tensor parallelism usually divides the same calculation evenly between matching GPUs. That does not work well here because the R9700 and Strix Halo...