NVIDIA DGX Spark Founders Edition Grace Blackwell Computer

The NVIDIA DGX Spark Founders Edition ships as a complete personal AI supercomputer: the GB10-based system, the 240W external power adapter, USB-C power cable, and NVIDIA's documentation and quickstart guide. For teams that need deeper deployment help, solutions are available from the EMARQUE AI team on request (pre-validation testing, setup assistance, EMARQUE AI stack, workflow tuning).

The Founders Edition is NVIDIA's own reference design for the DGX Spark family — the canonical implementation that partner variants (Gigabyte AI TOP ATOM, ASUS Ascent GX10, MSI EdgeXpert) all build to. Because all GB10-based DGX Spark units share identical silicon, the practical differences come down to thermal design, storage hardware, and chassis layout. Choosing the Founders Edition means going with NVIDIA's reference baseline — the spec other variants are measured against, with NVIDIA's own warranty and direct manufacturer support routing.

Published NVIDIA DGX Spark Founders Edition specifications: NVIDIA GB10 Grace Blackwell Superchip (20-core Arm CPU + Blackwell GPU); 128GB LPDDR5X unified memory; up to 4TB PCIe Gen5 NVMe storage; 10GbE networking plus NVIDIA ConnectX-7 SmartNIC for Spark Stacking; 3× USB 3.2 Gen 2×2 Type-C ports (data) and a separate Type-C power input; 240W external power adapter; NVIDIA reference chassis approximately 150mm × 150mm × 50.5mm at ~1.2kg.

Performance ceiling is approximately 1 petaFLOP of AI compute (FP4 sparse). This is NVIDIA's reference baseline — partner variants (Gigabyte AI TOP ATOM, ASUS Ascent GX10, MSI EdgeXpert) build on the same GB10 silicon with their own thermal tunings and storage choices, but the compute ceiling is shared. For exact stock-level config questions, email business@emarque.co.

The NVIDIA DGX Spark Founders Edition ships with NVIDIA's 3-year manufacturer warranty covering hardware faults on the unit and its components. As NVIDIA's own reference design, warranty is handled direct with NVIDIA — and EMARQUE's built-in role is local handling of that warranty:

• Local point of contact — pre-sale, post-sale, and warranty all routed through the same EMARQUE team. Real person, no ticket queue.
• RMA routing — if a hardware fault appears, we route the NVIDIA warranty claim on your behalf so you don't have to navigate the manufacturer's process yourself.

For deeper post-sale needs — pre-validation testing, OS reinstall, driver issues, network setup, Spark Stacking configuration, EMARQUE AI stack deployment — solutions are available from the EMARQUE AI team on request. Scope and charges quoted on enquiry.

NVIDIA's published performance for the DGX Spark platform (GB10 Grace Blackwell Superchip) — applicable to the Founders Edition as the reference baseline:

• ~1 petaFLOP of AI compute at FP4 sparse precision — Blackwell-generation tensor cores
• 128GB unified LPDDR5X memory at ~273 GB/s bandwidth — coherent across CPU and GPU (no PCIe copies, no host↔device transfers)
• Up to 4TB PCIe Gen5 NVMe — multi-GB/s storage I/O via GPU Direct Storage
• 200 Gb/s ConnectX-7 SmartNIC — required for Spark Stacking (linking two units to operate as a larger pool for 405B-class models or distributed training)

In practical workloads, the GB10 platform comfortably runs local LLM inference (Llama 3 family up to 70B at usable token rates, smaller models at production-grade throughput), local fine-tuning (LoRA / QLoRA on common open models), RAG pipelines with on-device vector search, and agentic-AI / multi-step inference chains. As NVIDIA's reference design, the Founders Edition establishes the baseline performance envelope — partner variants (Gigabyte, ASUS, MSI) may run slightly hotter or cooler depending on their thermal tunings, but the compute ceiling is shared across the DGX Spark family. For workload sizing relative to your specific use case, the EMARQUE AI team can review and quote.