Nvidia’s pint-sized powerhouse for AI development is here. Deliveries of the DGX Spark compact desktop have begun, and the launch made headlines when Jensen Huang personally handed one of the first units to Elon Musk. Built for researchers, startups, and creators who want to prototype and fine-tune models locally, this small form-factor system is designed to load large AI models into memory without the usual compromises.
The DGX Spark’s biggest advantage is its unified 128 GB LPDDR5x memory. Instead of juggling quantization tricks or shuttling data between system RAM and GPU VRAM, developers can pull entire large language models straight into memory, enabling smoother, offline experimentation and rapid iteration on the desktop.
Under the hood, the system pairs a 20-core Arm CPU—10 Cortex-X925 performance cores plus 10 Cortex-A725 efficiency cores—with 4 TB of NVMe M.2 storage. Nvidia rates the machine for up to one petaflop of Tensor performance in FP4 using sparsity, all while sipping just 240 watts. It runs DGX OS, a customized Ubuntu-based Linux environment tuned for AI workloads.
Despite its capability, this isn’t a raw speed replacement for high-end GPUs. Based on LMSys benchmarks, RTX 5080, RTX 5090, and Pro 6000 cards can deliver higher peak throughput. The key trade-off is memory architecture: those cards are blisteringly fast but limited by VRAM, whereas the DGX Spark’s unified 128 GB lets you keep massive models resident in memory for development convenience.
Desk-friendly dimensions complete the package. At just 150 x 150 x 50.5 mm and around 1.2 kg, the DGX Spark fits almost anywhere, bringing serious AI horsepower to labs and home offices alike.
Key specs at a glance:
– Up to 1 PFLOP Tensor performance (FP4 with sparsity)
– 20-core Arm CPU: 10 Cortex-X925 + 10 Cortex-A725
– 128 GB unified LPDDR5x memory
– 4 TB NVMe M.2 storage
– 240 W power draw
– DGX OS based on Ubuntu Linux
– Size: 150 x 150 x 50.5 mm; Weight: 1.2 kg
Price and availability:
– MSRP: $3,999.99
– General availability begins October 15, 2025
– Listed at Microcenter; not yet visible on Nvidia’s store on Amazon at the time of writing
Who should consider it:
– AI developers who need to load large models entirely into memory for local, offline prototyping and fine-tuning
– Teams prioritizing on-premises experimentation with private data
– Builders who value a compact, power-efficient desktop over maximum raw GPU throughput
If your workflow demands the absolute highest frame-by-frame compute, a flagship GPU may still be faster. But if you’re bottlenecked by memory limits when working with big models, the DGX Spark’s unified 128 GB approach could be the smarter, more streamlined tool for your desk.
