Inside the DGX Spark Cluster: Single-Node Configuration

The DGX Spark Cluster is a series of articles documenting each cluster size with real-world metrics, cost analysis, and operational details. This is the first article in the four-part series.

Read the full series:

• [Part 1] Single-Node Configuration — Baseline performance, memory headroom, and cost efficiency
• [Part 2] 2-Node Configuration — Distributed TP2 inference with connect
• [Part 3] 3-Node Configuration (Projected) — Scaling to TP3
• [Part 4] 4-Node Configuration (Projected) — Maximum deployment with HA storage

Overview

A single DGX Spark unit is a complete AI supercomputer in one compact form factor. Running the Qwen3.8-Flash-Next-FP8 model in tensor-parallel size 1 (TP1) on a single unit provides a baseline for comparison with the multi-node configurations.

Key takeaway: The single-node configuration uses all 128 GB of unified memory for model weights and KV cache, delivering strong per-request performance with no inter-node communication overhead.

Hardware Specifications

Software Stack

• vLLM: 0.30.0 (custom: dgx-spark/vllm:0.30.0-qwen4exp-fp8-ray)
• PyTorch: 2.13.0+cu130
• Transformers: 5.17.0
• Ray: Single-node cluster, 20 CPUs, 1 GPU reserved
• Kernel: 7.0.0-1019-nvidia
• NVIDIA Driver: 580.178.04

The Model: Qwen3.8-Flash-Next-FP8

• Type: 125B MoE (Mixture of Experts), 6B active parameters per token
• Checkpoint size: 173 GB — on a single node, the full model is loaded into 128 GB unified memory with TP1 (each of the 125B parameter shards fits on-chip via weight sharing)
• Native context: 262,144 tokens (256K)
• Hidden dimension: 2,560
• Layers: 48

vLLM Serving Configuration (TP1)

--tensor-parallel-size 1          # Single unit, no TP needed
--quantization fp8                # FP8 quantized checkpoint
--kv-cache-dtype bfloat16
--max-model-len 131072            # 128K combined input/output
--max-num-seqs 15                 # scheduler limit (practical: ~5-7 concurrent)
--max-num-batched-tokens 8192
--enable-chunked-prefill          # reduces KV cache memory waste
--gpu-memory-utilization 0.85
--limit-mm-per-prompt {"image":1,"video":0}
--reasoning-parser qwen3
--enable-auto-tool-choice
--tool-call-parser qwen3_coder

API endpoint: http://10.100.8.2:8000/v1
Health check: http://10.100.8.2:8000/health (returns 200)

Memory Analysis (Single Node)

With TP1, all model weights reside on the single GPU. Each Spark node (128 GB unified) allocates as follows:

Memory utilization sits at 93–95% on each node — consistent with the multi-node setup, but without the split-weight overhead.

Performance Metrics

GPU Utilization

• Average: 93–95% (each node)
• Peak: ~96%

Throughput

• Per-request: ~15 tokens/sec
• Estimated cluster total: ~167K output tokens/hour (half of the 2-node cluster)

Latency

Cost Analysis

Hardware Costs

Power Costs ($0.19/kWh)

Cost Per Token

• Per output token: ~$0.00013 ($0.13 per 1K tokens)
• Per training hour: ~$0.50–$1.00/hour

The single-node configuration offers the lowest per-unit cost at ~$4,700, making it the best entry point for the DGX Spark platform.

Advantages of Single-Node

• No interconnect latency — all communication is on-chip
• Simpler configuration — no Ray multi-node setup required
• Lower power per request — no switch or cable overhead
• Full 128 GB available — no memory split across nodes
• Easier to manage — single unit, single IP, single health endpoint

When to Use Each Configuration

Next in the Series

[Part 2] 2-Node Configuration — Adding a second Spark for TP2 inference. We cover the interconnect, distributed memory, and the jump to ~335K output tokens/hour.

About This Article

Based on actual DGX Spark cluster deployment data from October 2026. All metrics collected every 120 seconds via Hermes Agent v2.

Part 1: Single-Node Configuration · Part 2: 2-Node (Live) · Part 3: 3-Node (Projected) · Part 4: 4-Node (Projected)