DGX Spark guides
Forty guides on buying, running and operating a DGX Spark, from GPUwerk, which rents them out. They are grouped in four sections: buying and running a Spark, models and inference engines, clusters and scaling, and operations. Each card shows the guide's own summary.
Buying and running a Spark
Before you buy, then power, noise, setup, updates, storage and running costs.
What to know before buying a DGX Spark
An honest DGX Spark buying guide from a company that rents them out: real benchmarks, the 273 GB/s bandwidth ceiling, EU street prices, and buy vs rent.
DGX Spark power consumption and running cost
DGX Spark power draw, measured: 22-45 W idle, 60-90 W during LLM inference, about 160-200 W under full load, against a 240 W supply. Monthly cost at EU electricity prices.
DGX Spark noise: 29 dB at full load, 13 dB idle
NVIDIA declares the DGX Spark at 35 dB LWA and 29 dB LpA under maximum load, 19 and 13 dB at idle (ECMA-109). Reviewer readings, cooling and placement.
DGX Spark setup guide: first boot to a served model in one evening
A DGX Spark quick start from someone who racks them daily: DGX OS first boot, key-only SSH, choosing a serving stack, and serving your first model to your tools in one evening.
DGX Spark driver and firmware updates: what GPUwerk manages
What GPUwerk keeps updated at the host and node level on a rented DGX Spark, and what's inside your own root container and stays entirely your responsibility.
DGX Spark storage: what's persistent and what to back up
The 1 TB NVMe on a rented DGX Spark, exactly what survives a stop versus a terminate, and why keeping your own backup matters since GPUwerk keeps none.
Cold start vs always-on: choosing an instance uptime pattern
Stopping a DGX Spark between uses vs keeping it running around the clock: the cost, latency, and data tradeoffs, and how to decide which pattern fits actual usage.
Migrating an existing self-hosted AI deployment to a DGX Spark
A practical migration path from a workstation or another cloud GPU to a rented DGX Spark: re-quantizing for unified memory, moving data, and a cutover plan.
Cost forecasting and budgeting for AI infrastructure
How to estimate monthly AI infrastructure spend for self-hosting vs a metered API, with GPUwerk's $0.79/hour rate as a worked example and a budgeting checklist.
Total cost of ownership: GPU rental vs a cloud AI API
A per-token cloud AI API, a rented dedicated DGX Spark at $0.79/hour, and buying hardware outright, compared using only figures published on this site.
Models and inference engines
Which models to run, which engine to serve them with, and how to measure the result.
Best models for DGX Spark, October 2026 edition
Best models for DGX Spark, October 2026: Qwen3.6-35B-A3B NVFP4 reported at 86.3 tok/s on one Spark, DeepSeek-V4-Flash at about 44 tok/s on two. Verdict tables with sources.
llama.cpp vs vLLM vs SGLang on DGX Spark: tok/s and flags
vLLM 60.7 vs Ollama 42.1 tok/s for gpt-oss-120b on one DGX Spark, and 116 to 153 vs 42.5 at 8 users. Which engine for one user, a team or two Sparks, plus exact flags.
How to run a newly released model on a DGX Spark (GB10)
Run a new open-weight model on a DGX Spark in release week: memory fit, arm64 images, NVFP4 or GGUF, vLLM, SGLang or llama.cpp, and when it needs two Sparks.
Serve Qwen3 with vLLM on a DGX Spark
Install vLLM, serve Qwen3-32B-AWQ with an OpenAI-compatible API, connect Open WebUI, and keep it alive with systemd: a copy-paste guide for DGX Spark.
Quantized vs full-precision models in production
How to decide between a quantized and full-precision model for a production self-hosted deployment: memory fit, quality risk by task, and when to test both.
Fine-tune a model on your company's data, without your data leaving the building
Fine-tune an open LLM on company data without it leaving the building: LoRA and QLoRA memory tables, a 32B worked example, and fine-tune vs RAG.
Running embeddings and vector search on a DGX Spark
Run an embedding model and a vector database next to your generation model on one DGX Spark: memory budgeting, serving embeddings, and choosing a vector store.
Running vision and multimodal models on a DGX Spark
What's plausible on a DGX Spark's 128 GB of unified memory for vision-language models and image understanding, and why GPUwerk hasn't published dedicated VLM benchmarks yet.
ComfyUI and Stable Diffusion on a DGX Spark
Run ComfyUI, FLUX and SDXL on a DGX Spark: setup on ARM64, reported image times per model, unified-memory contention with LLMs, and thermal caveats.
Structured output and function calling on a self-hosted LLM
Setting up JSON-mode structured output and tool/function calling on self-hosted models with vLLM and llama.cpp, including which models support it and how to debug malformed calls.
Prompt caching to cut inference cost
How prefix and KV cache reuse cut repeated compute on a self-hosted deployment: what qualifies for a cache hit, how to structure prompts for it, and where the savings actually show up on an hourly-rented Spark.
Batch processing large inference jobs on a rented Spark
Run bulk document processing and dataset labeling on a rented DGX Spark: vLLM throughput settings, job sizing against the hourly rate, and checkpointing.
Evaluating model quality before switching models
Holdout test sets, A/B approaches, and regression checks for benchmarking a candidate model against your current one before switching in production.
How GPUwerk benchmarks the DGX Spark
What GPUwerk measures on a DGX Spark: single-stream versus concurrency, which engines and settings, and why the raw logs are published, not just a summary.
Clusters and scaling
Two Sparks, a fleet, several models on one machine, and sharing it with a team.
DGX Spark cluster: two Sparks, setup, limits and tok/s
Two DGX Sparks give 256 GB and, with tensor parallelism, up to about 1.8× decode (reported). Cable parts, ibdev2netdev port names, NVIDIA's playbook, models that need two Sparks, and tok/s.
Scaling from one DGX Spark to a fleet
When renting more single Sparks beats a two-node cluster, how to load balance across independent nodes, and the operational tradeoffs of running a small fleet.
GPU memory management for multiple models on one Spark
Budget 128 GB of unified memory across several models or containers on one DGX Spark, and know when to move to a two-node cluster or a second node instead.
GPU utilization monitoring and right-sizing your deployment
How to read actual GPU and memory utilization on a rented DGX Spark and use it to decide between staying put, scaling up to a two-node cluster, or scaling out to a fleet.
Running multiple containers and services on a DGX Spark
Docker container patterns for a single DGX Spark: when a plain docker-compose file is enough, and when the workload has outgrown it.
Updating or swapping models without downtime
A DGX Spark is a single machine, so true zero-downtime model updates need a second Spark to cut over to. Here's what that setup involves, and the honest alternative if you'll accept a short restart window instead.
Team and multi-user access to a shared DGX Spark
Patterns for sharing one rented DGX Spark across a team: separate Linux users with their own SSH keys, or a gateway in front so most people never need a shell.
Operations
Monitoring, backups, security and access control for a self-hosted deployment.
Monitoring and observability for a self-hosted LLM
What to watch on a single DGX Spark running an LLM, and which open-source tools do the job, from nvidia-smi to a Prometheus and Grafana stack. GPUwerk provides console and API access, not a monitoring product.
Backup and disaster recovery for your DGX Spark workload
GPUwerk keeps no copy of a rented DGX Spark's workspace. What to back up, where to send it, and how often to sync it.
Testing your disaster recovery plan for a self-hosted LLM
A backup you haven't restored from is a hypothesis. Restore drills, game days and RTO/RPO measurement for a self-hosted LLM on rented GPU hardware.
Security hardening a rented GPU instance
Practical in-instance hardening for a rented GPU: SSH key management, network exposure, running services without root, secrets, and ephemeral vs persistent state.
API key rotation best practices for a private LLM deployment
Rotating API keys and credentials for a self-hosted inference endpoint: scoping, rotation schedules, revocation, and what to automate versus do by hand.
Rate limiting and quotas for an internal LLM deployment
A raw DGX Spark has no per-user rate limiting. Add it yourself with LiteLLM's per-key rpm/tpm limits and budgets, or with a reverse proxy in front.
Audit logging for a compliance-sensitive self-hosted LLM deployment
Set up request and response audit logs in your own container on a rented DGX Spark: what to capture, where to store it, and where GPUwerk's job ends.
Setting up SSO and access control for a team LLM deployment
Putting an auth layer in front of a self-hosted inference endpoint so a team signs in with existing identities instead of sharing SSH keys or API tokens.
Model versioning and rollback strategy for a self-hosted deployment
Tracking which model weights and configs are actually running, and rolling back cleanly when a new version underperforms, on a self-hosted DGX Spark deployment.
For how the Spark compares with other machines, see the comparisons. For the hardware itself, see the specs page.