PNY NVDGXSPARK-PB NVIDIA DGX Spark, GB10 Grace Blackwell, 128GB LPDDR5x, 4TB NVMe M2 SSD, 10GbE, WiFi 7, Bluetooth 5.3
PNYPNY NVIDIA DGX Spark · GB10
A petaFLOP of AI.
On your desk.
PNY’s NVIDIA DGX Spark — the GB10 Grace Blackwell Superchip and 128 GB of unified memory in NVIDIA’s reference design. Prototype, fine-tune and run large models locally, then deploy to data center or cloud.
- GB10 Grace Blackwell
- 128 GB unified memory
- 1 PetaFLOP FP4
- DGX OS
The silicon
GB10 Grace Blackwell Superchip
A 20-core Arm CPU and a Blackwell GPU share one pool of memory over NVLink-C2C — no PCIe bottleneck between them. It is the architecture NVIDIA ships in DGX systems, on your desk.
- 20-core Arm CPU
- Blackwell GPU
- NVLink-C2C
Capacity
Room to run — and to scale
128 GB unified memory
Fine-tune models in the 30–70B range and run inference up to ~200B parameters — one pool, addressed by CPU and GPU alike, nothing to copy or partition.
Cluster when you outgrow one
The built-in NVIDIA ConnectX-7 NIC direct-connects two units — and scales to a four-unit cluster — when a single box isn’t enough.
Who it's for
Built for teams doing real AI work
Owned, local compute for the people who were paying for it by the hour.
Developers & ML researchers
Prototype and fine-tune in the 30–70B range; trade per-run cloud cost for a fixed, owned asset.
Regulated industries
Healthcare, finance and government teams that need on-prem, air-gapped-capable processing.
Research & academia
Capex-budget compute for imaging, genomics, astrophysics and coursework — no recurring cloud bill.
Robotics & edge vision
A local target for NVIDIA Isaac, Metropolis and Holoscan, from sensor processing to assistive tech.
Specifications
Technical specifications
- Superchip
- NVIDIA GB10 Grace Blackwell
- CPU
- 20-core Arm (10× Cortex-X925 + 10× Cortex-A725)
- GPU
- NVIDIA Blackwell — 5th-gen Tensor Cores, FP4
- AI performance
- Up to 1 petaFLOP (FP4)
- Unified memory
- 128 GB LPDDR5X — 256-bit, 273 GB/s
- Storage
- 4 TB NVMe M.2
- Networking
- NVIDIA ConnectX-7 (dual 200 Gbps QSFP) + 10 GbE
- Wireless
- Wi-Fi 7 · Bluetooth 5.3
- Rear I/O
- 4× USB-C · HDMI 2.1a · RJ-45 10 GbE
- Power
- 240 W external adapter (~170 W typical)
- Operating system
- NVIDIA DGX OS (Ubuntu Linux)
- Dimensions
- 150 × 150 × 50.5 mm · 1.2 kg
Questions
Frequently asked questions
Who is the PNY DGX Spark for?
AI developers, ML researchers, regulated-industry data scientists and research labs who want persistent, local AI compute — for fine-tuning in the 30–70B range and inference up to ~200B parameters.
Can it train large models from scratch?
It’s a fine-tuning and inference machine, not a from-scratch training cluster. For pretraining above ~70B parameters, look at NVIDIA DGX Station or a cloud cluster — we can advise on the right step up.
How does it compare to cloud GPUs on cost?
Cloud GPU sessions run roughly $2–$15 each and never stop. This turns that recurring opex into a fixed, owned asset — and keeps sensitive data on-prem.
What comes in the box?
The DGX Spark itself — GB10 Superchip, 128 GB unified memory, 4 TB NVMe, ConnectX-7 and NVIDIA DGX OS with the AI stack preloaded, plus the power adapter. Ask us to confirm exact in-box contents.
Is this the NVIDIA reference DGX Spark?
Yes — PNY ships NVIDIA’s DGX Spark reference design: the GB10 Superchip, 128 GB unified memory and the standard 150 × 150 × 50.5 mm chassis, backed by PNY as an NVIDIA elite partner.
What AI applications does it support?
Generative AI (fine-tuning and running LLMs), computer vision, predictive analytics, and simulation and research — a versatile local AI machine across industries.