NVIDIA Announces DGX Spark 64GB: 1-PetaFLOP AI Desktop
NVIDIA launches DGX Spark 64GB desktop featuring Grace Blackwell superchip, delivering 1 PetaFLOP of AI compute for local agents.
Stock photo for illustration only, not from the actual event
- NVIDIA introduces DGX Spark 64GB with GB10 Grace Blackwell superchip and unified memory.
- Delivers up to 1 PetaFLOP of AI compute power for running local agents and fine-tuning.
- Supports ConnectX-7 networking to cluster multiple units for expanded memory and compute.
- Ships with DGX OS ready on first boot, supporting PyTorch, Jupyter, and Ollama.
NVIDIA has announced the DGX Spark 64GB, a powerful desktop workstation designed to let developers run open-source models and always-on AI agents directly on their desks, bypassing metered cloud APIs whose token consumption has surged dramatically since early 2026.
The new hardware retains the same GB10 Grace Blackwell superchip, NVIDIA CUDA-accelerated AI software stack, and ConnectX-7 networking as the original model, while updating to 64GB of unified LPDDR5x memory instead of 128GB. NVIDIA positions this capacity as sufficient for today's leading 30–35B class open models, with the option to cluster multiple 64GB units together as workloads scale.
Stock photo for illustration only, not from the actual event
A key architectural choice is unified memory, allowing the 20-core Grace Arm CPU and Blackwell GPU to share a single memory pool over NVLink-C2C at 5x the bandwidth of PCIe Gen 5. This eliminates weight copying between system RAM and VRAM, ensuring models, KV caches, and tool processes reside in one address space.
"Run open models and always-on agents on your own desk, then cluster DGX Spark system as the work grows – instead of on a metered API."
NVIDIA
The introduction of the DGX Spark 64GB highlights a major shift toward local AI infrastructure, allowing developers and enterprises to maintain complete data privacy for sensitive code and documents without relying on expensive cloud services. Eliminating hardware bottlenecks through unified memory significantly enhances the performance of complex multi-step agent workloads.
On the software side, DGX OS ships ready on first boot with the NVIDIA AI stack, including PyTorch, Jupyter, and Ollama. Tools like NVIDIA NemoClaw and NVIDIA OpenShell provide built-in security and policy guardrails. The system operates on a standard wall outlet, requiring no dedicated server room or specialized cooling.
Source: MarkTechPost
Found something wrong in this article? Report an issue with this article
Comments
Leave a Comment