CustomRigs UAE is a trading brand operated by AMZ GLOBAL TRADING F.Z.E. — Ajman Free Zone License No. 43923 (FZA, United Arab Emirates). Registered address: Ajman Free Zone C1 Building, Ajman, UAE. Showroom (by appointment): Bur Dubai, Dubai, UAE. Contact: +971 56 468 9006 · info@customrigsuae.com. Not registered for VAT — prices are inclusive.

NVIDIA DGX Spark in UAE — 128GB Desktop AI Workstation

NVIDIA DGX Spark AI Workstation — 128GB, 4TB, Gold is available from CustomRigs UAE in Dubai. Approx. AED 27,000 — message us on WhatsApp for the latest price and stock, with UAE delivery, a 1-year warranty and a VAT invoice.

Why choose the NVIDIA DGX Spark for local AI?

Bring local AI development to your desk with the NVIDIA DGX Spark. This compact, complete workstation combines a NVIDIA GB10 Grace Blackwell Superchip, 128 GB of coherent unified system memory and a 4 TB NVMe SSD in a gold desktop enclosure. It is designed for AI developers, researchers, data scientists and teams building private local inference workflows.

The GB10 pairs a 20-core Arm CPU (10 Cortex-X925 and 10 Cortex-A725 cores) with a Blackwell GPU featuring fifth-generation Tensor Cores and fourth-generation RT Cores. NVIDIA specifies up to 1 petaFLOP of AI tensor performance at FP4 precision. This is a peak low-precision AI figure, not a claim of 1 PFLOP at FP32 or FP64, and actual throughput depends on the model, software and workload.

The CPU and GPU share 128 GB of coherent LPDDR5x memory over a 256-bit interface with 273 GB/s bandwidth. This is unified system memory, not 128 GB of separate dedicated graphics VRAM. NVIDIA describes inference support for models up to 200 billion parameters on one system, subject to quantization, model architecture and runtime memory overhead. Suitable workflows include local LLM inference, retrieval-augmented generation, AI prototyping, data science and supported fine-tuning workloads.

ConnectX-7 networking at up to 200 Gbps allows two DGX Spark systems to be linked for supported distributed workloads, including inference with models up to 405 billion parameters. A second system and suitable networking accessories are required and sold separately. Model limits are workload-dependent; they are not guaranteed training capacities. This listing is for one desktop DGX Spark, not a DGX rack server or a DGX SuperPOD installation.

Highlights

  • NVIDIA GB10 Grace Blackwell Superchip with 20-core Arm CPU
  • Blackwell GPU with fifth-generation Tensor Cores; up to 1 PFLOP at FP4
  • 128 GB coherent unified LPDDR5x memory; 273 GB/s bandwidth
  • 4 TB self-encrypting NVMe M.2 SSD with NVIDIA DGX OS
  • Local inference up to 200B parameters, subject to model and quantization
  • Connect two systems for supported workloads up to 405B parameters

Specifications

  • Superchip: NVIDIA GB10 Grace Blackwell
  • CPU: 20-core Arm: 10 × Cortex-X925 + 10 × Cortex-A725
  • GPU: NVIDIA Blackwell Architecture
  • Peak AI tensor performance: Up to 1 PFLOP at FP4 precision
  • System memory: 128 GB LPDDR5x, coherent unified CPU/GPU memory
  • Memory bandwidth: 273 GB/s
  • Storage: 4 TB NVMe M.2 SSD with self-encryption
  • High-speed networking: NVIDIA ConnectX-7, up to 200 Gbps
  • Operating system: NVIDIA DGX OS (Linux-based, Arm64)
  • Dimensions: 150 × 150 × 50.5 mm
  • Weight: 1.2 kg
  • Model / MPN: DGX Spark / DGXSPARKFOUNEDITUK
  • Architecture: NVIDIA Grace Blackwell
  • Tensor cores: 5th generation
  • RT cores: 4th generation
  • Memory interface: 256-bit
  • Single-system model support: Inference up to 200B parameters; model/quantization dependent
  • Two-system model support: Supported distributed inference up to 405B parameters; second system sold separately
  • USB: 4 × USB Type-C
  • Ethernet: 1 × RJ-45, 10 GbE
  • Wireless: Wi-Fi 7; Bluetooth 5.4
  • Display: 1 × HDMI 2.1a; DisplayPort over compatible USB-C (DP Alt Mode)
  • Audio: HDMI multichannel audio output
  • NVENC / NVDEC: 1 × / 1 ×
  • Power supply: 240 W
  • GB10 TDP: 140 W (CPU and GPU chip)
  • Finish: Gold

Frequently asked questions

What is NVIDIA DGX Spark designed for?

DGX Spark is a complete desktop AI workstation for developers, researchers and data scientists. Its GB10 Grace Blackwell Superchip supports local AI inference, prototyping, data science and compatible fine-tuning workflows. It is not a conventional Windows gaming PC or a DGX rack server.

Does it have 128 GB of dedicated GPU VRAM?

No. It has 128 GB of coherent unified LPDDR5x system memory shared by the CPU and GPU, with a 256-bit interface and 273 GB/s bandwidth. The 4 TB NVMe SSD is separate storage.

Can one DGX Spark run a 200-billion-parameter model?

NVIDIA describes inference support for models up to 200 billion parameters on one system. Actual fit and speed depend on quantization, model architecture, context length and software memory overhead. This is not a guarantee of full-precision inference or training a 200B model.

Can I connect two DGX Spark systems?

Yes. ConnectX-7 networking supports linking two systems for compatible distributed workloads, including inference with models up to 405 billion parameters under appropriate configurations. A second system and suitable interconnect accessories are sold separately; this listing includes one workstation.

What does the 1 petaFLOP performance rating mean?

It is NVIDIA’s peak AI tensor performance figure at FP4 precision. It does not represent FP32 or FP64 performance, and real application throughput depends on the workload and software.

Which operating system and software does it use?

It runs Linux-based NVIDIA DGX OS on a 20-core Arm CPU. Use compatible Arm64 applications, containers and AI frameworks. Standard x86 Windows software cannot be assumed to run unchanged; check your software requirements before ordering.

Can I upgrade the unified memory?

This listing is for the fixed 128 GB unified-memory configuration. It is not a conventional desktop with replaceable DIMM RAM, so plan your model and workload requirements around its configured capacity.

Why is the price approximate and how do I order in the UAE?

AED 27,000 is a guide price. Online purchase is disabled because current pricing and stock must be confirmed. Use Ask price on WhatsApp to receive the latest quotation, availability and lead time for model DGXSPARKFOUNEDITUK.

What is included and what warranty applies?

The quoted configuration is one gold NVIDIA DGX Spark with 128 GB unified memory, a 4 TB SSD and NVIDIA DGX OS. A monitor, keyboard, mouse, second workstation and interconnect accessories are not included unless stated in your quotation. The listing includes a one-year warranty; UAE delivery timing and regional package contents are confirmed when ordering.

How much does the NVIDIA DGX Spark cost in the UAE?

The NVIDIA DGX Spark is approximately AED 27,000 in the UAE. Pricing and stock change often, so CustomRigs UAE confirms the current price and lead time on WhatsApp before you order.

Where can I buy the NVIDIA DGX Spark in Dubai?

You can order the DGX Spark from CustomRigs UAE in Dubai, with delivery to Abu Dhabi, Sharjah and all emirates. Message us on WhatsApp to confirm price and availability.

Does the DGX Spark run Windows or games?

No. DGX Spark uses an Arm CPU and runs NVIDIA DGX OS, a Linux-based system for AI development. It is not designed for Windows software or gaming; for those, a custom RTX gaming PC or workstation is the better choice.

DGX Spark or Mac Studio for local LLMs?

Both use large unified memory. DGX Spark runs NVIDIA CUDA and the NVIDIA AI software stack natively, which most AI frameworks, fine-tuning tools and containers are built for. A Mac Studio suits teams already working in macOS. If your workflow depends on CUDA, DGX Spark is the more direct fit.

Is the DGX Spark worth it?

It is worth it if you need to run or fine-tune large models locally and privately with NVIDIA software, for example for LLM inference, RAG or prototyping before scaling to the cloud. For rendering, CAD or video work, an RTX PRO workstation usually gives better value.

Ask for the latest price on WhatsApp or browse NVIDIA DGX workstations and all workstations.