DGX Spark
#10 in KI-Trainings-Hardwarenvidia · siet 2025-10-15 · 2× · tolest 30. Juni 2026
16
Momentum
The NVIDIA DGX Spark is a compact desktop AI computer (150 × 150 × 50.5 mm) built around the NVIDIA GB10 Grace Blackwell Superchip, which pairs a Blackwell GPU (5th-gen Tensor Cores) with a 20-core Arm CPU via NVLink-C2C. The system features 128 GB of unified LPDDR5x memory (273 GB/s bandwidth), up to 1 petaFLOP of AI compute at FP4 precision (with sparsity), and a maximum system power draw of 240 W. It is primarily aimed at AI developers, researchers, and data scientists who want to run fine-tuning and inference locally without relying on cloud infrastructure.
Momentum-Verloop
04.04.03.07.
Features
| GPU/TPU Count | 1 integrated Blackwell GPU (GB10 SoC) with 6,144 CUDA cores, 5th-gen Tensor Cores, 4th-gen RT Cores |
| Max Model Size (Parameters) | Up to 200B parameters (inference/validation, single system); up to 405B parameters when clustering two DGX Spark systems via ConnectX-7 |
| Price Tier | From $3,999 (Founder's Edition, market launch October 2025); currently $4,699 on the NVIDIA Marketplace (as of 2026) |
| Compute Performance (TFLOPS/Exaflops) | 1,000 TFLOPS / 1 PetaFLOP at FP4 (with sparsity); ~500 TFLOPS FP4 without sparsity; ~31 TFLOPS FP32 |
| Storage Capacity (GB) | 128 GB unified LPDDR5x (CPU and GPU share the same memory pool, 273 GB/s bandwidth) |
