NVIDIA ships DGX Spark personal AI supercomputer with Blackwell
Listen to Briefing
0:00/1:41

What happened.NVIDIA began customer shipments of the DGX Spark personal AI supercomputer, delivering 1.2 petaflops of mixed-precision computing in a compact desktop form factor powered by Blackwell silicon.
Why it matters.Developing autonomous agent swarms requires rapid trial-and-error cycles that strain cloud budget limits and expose pre-release code to remote networks. Bringing liquid-cooled Blackwell workstations into private engineering environments allows builders to fine-tune local reasoning models without recurring cloud compute bills.
The technical specifications:
- Compute architecture: Packs two Blackwell Ultra GPUs alongside a Grace CPU to deliver 1.2 petaflops of FP4 inference throughput within a quiet workstation chassis.
- Unified memory: Provides 288GB of unified HBM3e memory with 8 terabytes per second bandwidth to run 70-billion-parameter models locally without quantizing weights.
- Thermal engineering: Uses a closed-loop liquid cooling system that limits acoustic output to 34 decibels under continuous peak computational load.
- Developer stack: Pre-loaded with NVIDIA AI Enterprise, NeMo framework, and CUDA-X libraries for local agent fine-tuning and retrieval workflows.