Desktop AI supercomputing: 1 PFLOP FP4 performance fits in your palm.
The NVIDIA DGX Spark GB10 packs Grace Blackwell architecture into a 150 mm chassis, delivering 1 PFLOP FP4 tensor performance and 128 GB LPDDR5x memory. A 4 TB NVMe workstation that puts AI supercomputing on your desktop at just 240W.
1 PFLOP FP4 Tensor Power
Grace Blackwell architecture delivers petaflop-scale AI compute for training, fine-tuning, and inference.
128GB Unified Memory
LPDDR5x at 273 GB/s feeds massive models without data movement bottlen
NVIDIA DGX Spark GB10 Grace Blackwell AI Supercomputer 128GB 1 PFLOP FP4 4TB
Engineered for the AI frontier
The DGX Spark GB10 packs a 20-core Arm CPU cluster—10 Cortex-X925 and 10 Cortex-A725 cores—alongside Blackwell-class Tensor Cores into a 150 mm square chassis. With 4TB NVMe M.2 storage, ConnectX-7 networking, Wi‑Fi 7E, and full DGX OS compatibility, it bridges from prototype to production without leaving your desk. At just 240W, it delivers data center-class AI acceleration in a compact, quiet desktop form factor.
Specifications
Performance
Suite
1 PFLOP
FP4 Tensor Performance
The Full Picture.
A Revolutionary Desktop AI Workstation
The NVIDIA DGX Spark GB10 Grace Blackwell AI Supercomputer transforms your desktop into an artificial intelligence computing hub. With 128 GB of LPDDR5x memory and a bandwidth of 273 GB/s, it effortlessly runs large language models and the most resource-intensive neural networks.
Data center-class performance
Thanks to the NVIDIA Grace Blackwell architecture and its 20 optimized Arm cores, this workstation delivers 1 PFLOP in FP4. AI researchers, developers, and generative content creators benefit from accelerated training and fine-tuning, right at their workstations.
Storage and connectivity designed for the workflow
- 4 TB NVMe M.2: large datasets and model checkpoints are always instantly accessible
- ConnectX-7, Wi-Fi 7E, Bluetooth: Seamless integration into existing network infrastructures
- HDMI and USB-C: Multi-monitor setup for viewing results in real time
Power-packed, flexible placement
With compact dimensions of 150 x 150 x 50.5 mm and low power consumption of 240 W, this personal supercomputer can be installed anywhere. Compatible with NVIDIA DGX OS, it integrates seamlessly into professional AI development ecosystems.
Built by Enthusiasts.
I ran a 70B parameter fine-tuning job overnight on my desk. What used to require cloud rental now sits beside my monitor, silent and self-contained. The unified memory architecture eliminated the sharding headaches entirely.
As a boutique integrator, I needed a demonstrable edge for client pitches. The DGX Spark lets me bring live inference demos to boardrooms—no internet, no cloud credits, no latency. The ConnectX-7 throughput still shocks
