DeepSeek V4 Flash on Two DGX Sparks

Measured / sourced results

What the numbers mean

What the numbers mean guidance: Public two-node results can exceed one-node throughput, but scaling may be sublinear or even regress for a communication-heavy configuration. Expect the main win to be fit and headroom; treat speedup as a measured bonus. Choose two Sparks when one node cannot hold the reviewed checkpoint with 20% headroom. If one node already fits, add the second only after a measured concurrency or TTFT bottleneck.

What stands out

Primary benefit: 256 GB aggregate capacity.

Do not assume 2× speed; verify communication overhead with your batch and context.

Use one larger professional server when operational simplicity outweighs compactness.

Evidence quality / source

Evidence quality / source guidance: NVIDIA documents Spark interconnect workflows; DeepSeek documents the model; public two-node results are examples rather than guaranteed scaling factors. Aggregate memory is useful only if the runtime can shard the exact model architecture and quantization.

What is unknown

• Aggregate memory is useful only if the runtime can shard the exact model architecture and quantization.
• A benchmark over a particular link and batch size should not be treated as a universal two-node multiplier.

Workload fit

Workload fit guidance: Recommended for researchers who need this exact model locally and can manage distributed serving. It is not the simplest choice for a small team that just needs a fast private assistant. Two systems expose 256 GB of aggregate memory, enough for mixed FP4/FP8 weights plus useful context if the runtime shards experts and cache correctly. Keep at least 25–40 GB aggregate free for communication buffers and runtime workspaces.

Alternatives

Alternatives guidance: A 192 GB professional GPU server reduces networking complexity. A smaller MoE such as Mistral Small 4 or Qwen3.6 often gives better interactive value on one Spark.

Planner CTA

Compare this recommendation against your model, context, concurrency, latency, and budget in the ComputeSage Planner.

Sources / methodology

Recommendations combine official specifications, public model/runtime documentation, adjacent benchmark observations, and clearly labeled engineering estimates. Review the ComputeSage evidence methodology.

• DeepSeek-V4-Flash model card
• NVIDIA DGX Spark User Guide
• vLLM installation and hardware documentation
• llama.cpp multi-GPU documentation
• MiaAI-Lab DeepSeek V4 Flash 0731 dual-Spark benchmark matrix