NVIDIA DGX Spark™ – Personal AI Desktop Supercomputer – Desktop GB10 Grace Blackwell Chip

The NVIDIA DGX Spark puts data-center-class AI compute in a chassis the size of a hardback book. It is built for machine learning engineers, researchers and small AI teams who need to prototype, fine-tune and run large models locally instead of queuing for cluster time or paying cloud GPU bills. If you spin up an A100 or H100 instance every time you test a fine-tune, this replaces that loop with hardware on your desk, with no idle-GPU billing and no dataset leaving the building.

Who it's for

• Fine-tuning an open-weight 70B model with LoRA/QLoRA
• Iterating on a RAG pipeline against a private document set, offline and air-gapped if required
• Benchmarking quantization and serving inference for models in the 70B–200B class

Key features

• NVIDIA GB10 Grace Blackwell Superchip: a 20-core Arm CPU (10x Cortex-X925 performance cores + 10x Cortex-A725 efficiency cores) fused to a Blackwell-architecture GPU with 5th-generation Tensor Cores and 4th-generation RT Cores over NVLink-C2C, rated for up to 1 petaFLOP of FP4 AI compute with sparsity.
• 128GB unified LPDDR5x memory at 273GB/s: CPU and GPU share one coherent 128GB pool over a 256-bit memory interface, with no PCIe round-trip copying between system RAM and VRAM. It holds models that would run out of memory on a 24GB or even 48GB discrete workstation GPU.
• Run and fine-tune models up to ~200B parameters: at FP4 precision, a single unit hosts models in the 70B–200B class for local inference, LoRA/QLoRA fine-tuning and RAG pipeline development.
• 4TB self-encrypting NVMe SSD: local headroom for multiple model checkpoints, quantized variants, container images and full training datasets.
• ConnectX-7 Smart NIC (200Gbps): pair two DGX Spark units with a certified QSFP112 interconnect cable to pool 256GB of unified memory and run inference on models up to roughly 405B parameters.
• Full NVIDIA DGX OS + AI stack, pre-installed: CUDA, cuDNN, NCCL, PyTorch and TensorFlow containers, NVIDIA AI Workbench, and accelerated frameworks such as Isaac, Metropolis and Holoscan. Workloads are container-portable to other Blackwell systems or DGX Cloud without code changes.
• Desktop footprint: 150 x 150 x 50.5mm and 1.2kg, powered from a single 240W supply, with 1x HDMI 2.1a and up to 3x DisplayPort 2.1 over USB-C.
• Connectivity: 4x USB Type-C, 10 Gigabit Ethernet, Wi-Fi 7, Bluetooth 5.4.

Specifications

Good to know

• The 1 petaFLOP figure is a peak FP4-with-sparsity number. Real dense-precision throughput will be lower.
• It is not a from-scratch training rig: with 273GB/s memory bandwidth and single-GPU-equivalent compute, pre-training multi-billion-parameter models from random initialization is cluster or cloud territory.
• A 70B model in FP4/INT4 quantization, or a 30B model at higher precision with generous context length, loads without tensor-parallel sharding.
• Two configurations: a Single Unit, or a Dual Unit + Interconnect Cable bundle (two units plus an NVIDIA-certified ConnectX-7 QSFP112 cable) for 256GB of pooled memory.

Choose the single unit if your largest model fits in 128GB, or the dual-unit bundle for models up to roughly 405B parameters.