NVIDIA DGX Spark AI Supercomputer

NVIDIA DGX Spark is a compact desktop AI supercomputer for local AI development. It uses the NVIDIA GB10 Grace Blackwell Superchip with a 20-core Arm CPU, Blackwell GPU, 128GB unified memory, and NVIDIA's full AI software stack. Built for developers, researchers, AI teams, and businesses that want to prototype, fine-tune, and run AI models locally instead of relying on cloud GPUs.

DGX Spark suits AI developers, data scientists, research teams, software teams building AI agents, businesses needing on-premise AI, and universities or labs running local AI workloads. The common need: build, test, fine-tune, and run AI models locally instead of relying on cloud GPUs.

NVIDIA positions DGX Spark for models up to ~200B parameters. In real-world use, EMARQUE's testing and client deployments show ~120B parameters as the practical optimum on a single Spark — beyond that, throughput and latency degrade. For larger models, NVIDIA's Spark Stacking (multi-node clustering) links up to four units: two units reach ~405B-parameter models, and four units pool 512GB of unified memory for large mixture-of-experts models.

Single Spark fits well: 7B–14B (excellent), 30B–32B (suitable), 70B (workload-dependent), ~120B (real-world ceiling), ~200B (pushes limits), 405B+ (needs a 2–4 unit cluster or step-up).

Yes. LLMs: local inference, RAG workflows, AI assistants, coding models, fine-tuning small-to-mid LLMs. VLMs: vision-language models, document AI, multimodal workflows (heavier high-resolution VLMs may need RTX PRO 6000 or DGX Station). AI agents: NVIDIA-positioned platform for tool-calling, coding agents, private assistants, and long-running internal AI systems.

Fine-tuning: yes — small-to-mid model fine-tuning, LoRA / QLoRA, private data adaptation, prototyping before scaling. Training from scratch: only small experiments. Large foundation model training, heavy multi-user inference, or production AI serving need RTX PRO 6000 Blackwell, DGX Station GB300, or server-class systems.

Yes. DGX Spark works well as a private local AI server for internal teams, demos, RAG systems, coding assistants, and chatbot testing. A strong fit for businesses that want sensitive data — files, customer info, R&D — to stay on-premise instead of routing to public cloud AI services.

DGX Spark is built for AI development with NVIDIA's preconfigured AI software stack — not gaming. For gaming or general workstation use, choose an EMARQUE Gaming PC.

For AI workloads compared to a custom AI PC built around an RTX 5090 (32GB GDDR7):

• DGX Spark wins when models exceed ~32GB — its 128GB unified memory handles 70B–120B-class models that simply won't fit on a single RTX 5090, plus the AI stack is preconfigured out of the box.
• RTX 5090 wins for smaller models (≤30B) — about 6.6× higher memory bandwidth (1,792 GB/s vs Spark's 273 GB/s) and FP8 / FP4 raw throughput give faster token generation, plus broader workstation flexibility (rendering, gaming, mixed workloads).

Different sweet spots. EMARQUE can build either, and our solutions team will tell you honestly which fits your workload better.

DGX Spark is built for capacity, not peak speed. Its 128GB unified memory holds models a gaming GPU can't, but its memory bandwidth is modest at 273 GB/s (roughly 6.6× lower than an RTX 5090's 1,792 GB/s), and that sets the token-generation rate. From independent mid-2026 reviews (single-user; figures vary by model, quantisation and framework):

• 8B–20B models: interactive, from tens of tokens/sec single-user to hundreds batched.
• 70B models: comfortably usable, roughly 35–45 tokens/sec.
• 120B models (e.g. GPT-OSS 120B): around 38 tokens/sec. The value is that it runs locally at all, which a 32GB GPU can't do.

The platform keeps getting faster in software: NVIDIA's January 2026 (CES) update delivered up to 2.5× gains on some workloads, and the June 2026 update added up to 2.6× throughput on vLLM alongside multi-node clustering. For peak speed on models that fit in 32GB, an RTX 5090 or RTX PRO 6000 is faster; for running 70B–120B models locally, DGX Spark is the compact, efficient choice.

Yes. As of NVIDIA's June 2026 update you can cluster up to four DGX Spark units (called Spark Stacking, or multi-node clustering), not just two. Each unit has two ConnectX-7 200Gb/s ports, and NVIDIA's Cluster Assistant in NVIDIA Sync configures the cluster:

• 2 units: one QSFP cable, no switch. 256GB unified memory, models up to ~405B parameters.
• 3 units: direct ring topology using both ports, no switch. 384GB unified memory.
• 4 units: via a 200GbE QSFP switch. 512GB unified memory, enough for large mixture-of-experts models and multi-agent pipelines.

EMARQUE supplies and configures the full cluster (extra units, switch, cabling and Cluster Assistant setup), so you get a working multi-node system rather than a box of parts. Need more than four units, or a single larger machine? Step up to DGX Station GB300, RTX PRO 6000 Blackwell, or a multi-GPU AI server.

Two natural step-ups beyond DGX Spark:

RTX PRO 6000 Blackwell Workstation — when you need stronger raw GPU performance plus professional workstation flexibility (rendering, simulation, visualization). Best for mixed AI + creator workloads.

DGX Station GB300 — when DGX Spark is too small and a workstation isn't enough. Larger memory, far more compute, trillion-parameter models, multi-user workloads. The right step for serious on-prem AI infrastructure.

EMARQUE AI helps you size it. Contact business@emarque.co.

• DGX Spark — local AI development, LLM/VLM testing, RAG, fine-tuning small-to-mid models
• 2–4× DGX Spark cluster — larger models via pooled memory (~405B on two units, 512GB on four), distributed local AI
• RTX PRO 6000 Blackwell — high-speed AI plus rendering, simulation, mixed creator workloads
• DGX Station GB300 — serious on-prem AI, 1T-parameter models, multi-user, advanced training
• Gaming? — EMARQUE Gaming PC, not DGX Spark

Yes. EMARQUE has helped 300+ Malaysian clients integrate AI and high-performance computing — enterprises, SMEs, research labs, universities, government agencies, and creative studios. Deployments span private LLM stacks, on-prem RAG, document AI, computer vision, and custom AI workstations. Common thread: local data control, predictable performance, Malaysia-based support. Doing this since 2016.

DGX Spark ships with NVIDIA's full DGX OS + AI Enterprise software stack — CUDA, container runtimes, NIM microservices, AI Workbench, and pre-built framework images (PyTorch, vLLM, etc.). You can run models out of the box.

Every DGX Spark from EMARQUE includes a free EMARQUE AI Utility, installed on request, so your team is productive from day one instead of spending weeks on setup. For a fuller deployment, the EMARQUE AI team also builds a complete local AI stack as a solution on request: pre-configured chat UI, private RAG pipeline with vector search, document ingestion, and model serving. Scope and charges for the full stack are quoted on enquiry.

Beyond software, EMARQUE is a DGX Spark expert and systems integrator in Malaysia — not just a reseller. EMARQUE's solutions team handles sizing and configuration; networking (QSFP / ConnectX-7); private LLM/VLM/RAG and AI-agent deployments; team training; and Malaysia-based post-sale support — all available as EMARQUE AI solutions on request.

On-prem AI is the right choice when privacy (data sovereignty, PDPA compliance, sensitive records), predictable cost (vs. unbounded cloud GPU bills), latency (no cloud round-trip), or customization (full control of models and stack) matters more than cloud convenience. EMARQUE designs DGX Spark deployments for these realities — for many Malaysian businesses, on-prem with EMARQUE is the better long-term fit.

Yes. EMARQUE AI is our dedicated AI compute division — full NVIDIA AI hardware stack and on-prem deployment for Malaysian businesses, research labs, and enterprises: NVIDIA Personal AI (DGX Spark, 2× Spark), AI workstations (RTX PRO 6000 Blackwell and others), DGX Station GB300, GPU servers, plus full integration support. Visit EMARQUE.AI or contact business@emarque.co.