I Ran DeepSeek V4 Flash Across Two DGX Sparks Over Ethernet

Developer Kevin Tang successfully demonstrated that two NVIDIA DGX Spark units can run the DeepSeek V4 Flash model using a standard 10Gb Ethernet connection instead of the recommended 200Gb fabric. By configuring NCCL to use TCP sockets over the direct Ethernet link, Tang achieved stable tensor-parallel inference. While the improvised setup resulted in roughly 40-46% of the performance compared to the official 200Gb/s specification, it proved functional for both interactive chat and batch processing. The experiment highlights that while high-speed interconnects are critical for latency-sensitive applications, standard networking can suffice for capacity-constrained tasks. Tang's setup allowed the 284-billion parameter model to operate across the combined 256GB of unified memory, providing a viable, albeit slower, alternative for users awaiting specialized cabling. This approach serves as a practical proof-of-concept for distributed LLM inference on hardware clusters.