NVIDIA GB10 Grace Blackwell Superchip

A personal AI supercomputer on your desk.

DGX Spark packs a petaFLOP of AI performance and 128 GB of unified memory into a box roughly the size of a hardcover book. Prototype, fine-tune, and run 200-billion-parameter models locally, then move to the data center or DGX Cloud without changing a line.

View specifications What you can run

0PFLOP

FP4 AI performance

0GB

Unified memory

Parameters, local

0mm

Square footprint

FIG. 01 — GB10 desktop unit, approx. 1.2 kg

01 / OVERVIEW

Data-center architecture, shrunk to a 150 mm square.

DGX Spark is built on the same Grace Blackwell design language as NVIDIA's rack-scale systems. The GB10 Superchip joins a Blackwell GPU and a 20-core Arm CPU over NVLink-C2C, so the whole 128 GB memory pool is visible to both sides without copies across a PCIe bus.

Compute

Blackwell GPU with fifth-gen Tensor Cores

Delivers up to 1 petaFLOP of FP4 AI performance for training, fine-tuning, and inference on transformer models, with native support for the FP4 precision that Blackwell introduced.

Memory

128 GB of coherent LPDDR5x

CPU and GPU share one unified address space. Large models load into a single pool of memory instead of being split across a host and a small VRAM buffer.

Platform

Runs the full NVIDIA AI software stack

Ships with DGX OS, an Ubuntu-based Linux distribution preloaded with CUDA, drivers, containers, and the NIM microservice catalog. Your workflow on Spark is the same one you would run on DGX Cloud.

02 / SPECIFICATIONS

Technical specifications.

Published figures for the production DGX Spark unit. Peak AI performance is quoted at FP4 precision with sparsity, per NVIDIA's convention for Blackwell systems.

DGX Spark key specifications
Component	Specification	Unit	Notes
Superchip	NVIDIA GB10 Grace Blackwell	—	CPU and GPU connected via NVLink-C2C interconnect
CPU	20-core Arm	cores	10 × Cortex-X925 performance + 10 × Cortex-A725 efficiency
GPU	NVIDIA Blackwell architecture	—	Fifth-generation Tensor Cores, fourth-generation RT Cores
AI performance	1	PFLOP	FP4 with sparsity, peak
System memory	128	GB	LPDDR5x, unified and coherent across CPU and GPU
Memory bandwidth	273	GB/s	256-bit interface
Storage	1 or 4	TB	NVMe M.2 with self-encryption
Networking	NVIDIA ConnectX-7	—	Smart NIC for high-speed clustering of two units; plus 10 GbE, Wi-Fi 7, Bluetooth 5.4
I/O	4 × USB-C, HDMI 2.1a	—	Display output up to 4K
Model capacity	200	B params	Local inference on a single unit; ~405B with two units linked
Dimensions	150 × 150 × 50.5	mm	Approximately 1.2 kg
Operating system	NVIDIA DGX OS	—	Ubuntu-based Linux with the NVIDIA AI software stack preinstalled

* Specifications summarized from NVIDIA's public product announcements. Verify current figures at nvidia.com before purchase.

03 / WORKLOADS

What you can run.

The 128 GB memory pool is the point. Workloads that used to require a rented multi-GPU node, or that could not run at all on a consumer card, fit on a single desk.

WL-01TRAIN

Fine-tune mid-size models

Run LoRA and full fine-tuning on models up to about 70B parameters directly on the unit, using PyTorch, NeMo, or Hugging Face tooling as-is.

Fine-tuning

WL-02INFER

Local inference at 200B

Serve open-weight models up to roughly 200B parameters on one unit with NIM microservices, vLLM, or llama.cpp, with the data never leaving the room.

Inference

WL-03SHIP

Prototype, then deploy

Because Spark runs the same containers and drivers as DGX Cloud and DGX systems, a pipeline built on the desk can be moved to the data center for production with minimal changes.

DGX Cloud

WL-04SIM

Robotics and edge simulation

Develop vision-language-action models, run Isaac and Omniverse simulation workloads, and iterate on edge inference before pushing to Jetson devices.

Physical AI

04 / SOFTWARE STACK

Everything installed. Nothing to configure.

DGX OS arrives with the NVIDIA AI platform preloaded. Open a terminal, pull a NIM container, and you are serving a model. The stack below is what you get on first boot.

01
CUDA-X librariesCUDA, cuDNN, TensorRT, and the full accelerated library set.
02
NVIDIA NIM microservicesOptimized, containerized inference endpoints for popular open models.
03
FrameworksPyTorch, TensorFlow, and JAX builds tuned for Blackwell.
04
NeMo and AI BlueprintsReference workflows for RAG, agents, and speech.
05
Containers and toolingDocker, NGC catalog access, and Jupyter ready to go.

spark@dgx:~$ cat /etc/dgx-release | head -2
DGX_NAME="DGX OS"
DGX_PLATFORM="DGX Spark"

spark@dgx:~$ nvidia-smi --query-gpu=name,memory.total --format=csv
name, memory.total [MiB]
NVIDIA GB10, 131072 MiB

spark@dgx:~$ nvcc --version | tail -1
Cuda compilation tools, release 13.x

spark@dgx:~$ python -c "import torch; print(torch.cuda.get_device_name(0))"
NVIDIA GB10

spark@dgx:~$ docker run --gpus all nvcr.io/nim/meta/llama-3.1-70b-instruct
Pulling image ...................................... done
Loading weights into unified memory ................ done
Uvicorn running on http://0.0.0.0:8000

spark@dgx:~$

05 / WHY LOCAL

The case for compute you can touch.

Cloud GPUs are the right answer for scale. For the daily loop of experiments, evaluations, and sensitive data, a machine on the desk removes three recurring costs.

Privacy by default

Proprietary code, customer records, and unreleased research stay on hardware you control. Nothing is uploaded to run a prompt.

CLOUD API → data leaves the buildingDGX SPARK → data stays on the desk

Latency you can feel

No queueing for a shared node, no cold starts, no round trip to a region. Iteration speed is limited by the model, not the network.

CLOUD API → queue + network + cold startDGX SPARK → first token from the desk

Cost predictability

One purchase, then unlimited runs. Overnight evaluation sweeps and long fine-tunes stop showing up as a surprise on a monthly invoice.

CLOUD API → per-token, per-hour billingDGX SPARK → fixed cost, unmetered use

06 / SCALE OUT

Two units. One 405B model.

Connect two DGX Spark systems over the integrated ConnectX-7 networking to pool 256 GB of unified memory.
Run inference on models up to roughly 405 billion parameters, including the largest open-weight Llama checkpoints.
Move the same containers to DGX Cloud or a DGX rack when the workload outgrows the desk.