Nvidia's DGX Spark AI mini-PC goes up for sale October 15 – 1 petaFLOP (www.tomshardware.com)

Nvidia has set an October 15 retail launch for the DGX Spark mini‑PC, available from Nvidia and partners (Dell, Asus, MSI, HP). The compact system pairs an Arm-based Grace Blackwell GB10 CPU with a Blackwell GPU and a unified 128 GB of LPDDR5X shared memory, delivering up to ~1 petaFLOP of AI inferencing throughput—contingent on FP4 quantization and model sparsity. A single unit claims support for models up to ~200 billion parameters (FP4), and two units can be linked via a built‑in ConnectX‑7 NIC to double memory and compute. It runs Nvidia’s DGX OS (Ubuntu fork) and full CUDA tooling; launch price is $3,999 (up from an initial $3,000 target). Why it matters: DGX Spark targets the growing need to run large‑scale models locally without multi‑rack servers, addressing the GPU‑local memory bottleneck that limits inference on consumer cards (e.g., 32 GB on RTX 5090). By combining large shared RAM, an Arm CPU optimized for AI, and native CUDA support plus integrations with partners like Hugging Face, Docker, Microsoft and Google, it promises a turnkey platform for developers and edge inference. Tradeoffs include reliance on aggressive quantization/sparsity for peak claims and a higher price relative to Ryzen AI‑based mini‑PCs that offer big memory but lack native CUDA compatibility.