Chapter 1 · Part I

Introduction: What You Are Building

The DGX Spark Bundle

The NVIDIA DGX Spark Bundle is two DGX Spark personal AI supercomputers sold together with everything needed to link them into a single distributed cluster.

ResourcePer SparkCombined
GPU1× NVIDIA GB10 Blackwell2× GPUs
GPU Memory128GB unified256GB unified
CPU Cores20 (ARM64)40 cores
InterconnectConnectX-7 RDMA (QSFP)
OSDGX OS (Ubuntu 24.04.4 LTS)

Software Stack

🐧
DGX OS
Ubuntu 24.04.4 LTS with NVIDIA-optimized kernel and CUDA 13.0
Kernel 6.17
☸️
k3s
Lightweight Kubernetes — full API, minimal footprint, containerd runtime
v1.35.5
🎮
GPU Operator
Automatically deploys device plugin, DCGM exporter, container toolkit
NVIDIA Helm
KubeRay
Kubernetes operator managing the distributed Ray cluster lifecycle
Ray 2.49.2
🤖
vLLM
High-performance LLM inference — PagedAttention, tensor parallelism, OpenAI API
0.10.1.1
🔀
AIBrix
AI inference gateway — routing, multi-tenancy, agent lifecycle management
v0.6.0

Cluster Architecture

Your Network 192.168.86.x Spark 1 — k3s Master 192.168.86.30 · spark-720e GB10 Blackwell GPU · rank 0 128 GB unified memory ● Ray Head Pod ● vLLM API Server :8000 ● KubeRay Operator ● AIBrix Control Plane ● Prometheus / Grafana ● NVIDIA GPU Operator Spark 2 — k3s Worker 192.168.86.26 · spark-7229 GB10 Blackwell GPU · rank 1 128 GB unified memory ● Ray Worker Pod ● Tensor Parallel rank 1 ● NCCL all-reduce ● k3s agent QSFP RDMA · ConnectX-7
Figure 1.1 — DGX Spark Bundle: two-node cluster topology

Time Estimates

ChapterEstimated Time
Hardware Setup60–90 min
CUDA Updates30–60 min
k3s + GPU Operator20–30 min
KubeRay15–20 min
vLLM (incl. model download)30–60 min
AIBrix10–15 min
Monitoring20–30 min
Total3–5 hours