NVIDIA DGX Spark is a compact desktop AI supercomputer powered by the Grace Blackwell GB10 superchip, designed to build and run autonomous agents locally. It delivers up to one petaflop of AI performance, features 128GB unified memory, and supports large models without relying on cloud resources. Integrated NVIDIA AI software, tools, and networking enable fast development, fine tuning, and inference for advanced workloads. Its energy efficient design supports always on operation for developers, researchers, and data scientists. DGX Spark accelerates prototyping, edge applications, and production grade AI with enterprise ready performance, security, and scalability in a desktop form factor compactly.
| OS | NVIDIA DGX™ OS |
| CPU | 20-core Arm, 10 Cortex-X925 + 10 Cortex-A725 Arm |
| GPU | Blackwell Architecture |
| NIC | ConnectX-7 NIC @ 200 Gbps |
| USB | 4x USB TypeC |
| Wi-Fi | WiFi 7 |
| Storage | 4 TB NVME.M2 with self-encryption |
| Ethernet | 1x RJ-45 connector 10 GbE |
| RT Cores | 4th Generation |
| Bluetooth | BT 5.4 |
| GB10 TDP2 | 140 W |
| CUDA Cores | Blackwell Generation |
| Architecture | NVIDIA Grace Blackwell |
| Audio-output | HDMI multichannel audio output |
| Power Supply | 240 Watts |
| Tensor Cores | 5th Generation |
| NVENC | NVDEC | 1x | 1x |
| System Memory | 128 GB LPDDR5x, coherent unified system memory |
| System Weight | 1.2 kg |
| Memory Bandwidth | 273 GB/s |
| Memory Interface | 256-bit |
| System Dimensions | 150 mm L x 150 mm W x 50.5 mm H |
| Display Connectors | 1x HDMI 2.1a, Up to 3x DisplayPort over USB-C (DP Alt Mode) |
| Tensor Performance1 | Up to 1 PFLOP FP4 |
