Grace meets Blackwell.
A 20-core Arm CPU and Blackwell GPU work together in the NVIDIA GB10 Superchip.
INTEGRATED BY DESIGNPersonal AI supercomputer
Big ideas.
Local compute.
Your next breakthrough starts here. Grace Blackwell performance. Room for ambitious models. All within arm’s reach.
Find your workload↗Inside the Spark ↓GB10 / GRACE BLACKWELL
150 × 150 × 50.5 mm
DESKTOP SCALE. EXCEPTIONAL POTENTIAL.1 PFLOP
UP TO · FP4 SPARSE COMPUTE
128 GB
COHERENT UNIFIED MEMORY
200 B
MODEL PARAMETERS · UP TO
4 TB
NVME STORAGE
01 / ENGINEERED FOR WHAT’S NEXT
Small footprint.
Serious architecture.
GPU. CPU. Memory. One tightly connected platform for the work you want to do next.
One pool. More room.
128 GB of coherent unified memory shared by CPU and GPU.
273 GB/S MEMORY BANDWIDTHBuilt to connect.
ConnectX-7 networking opens a path to workflows across multiple Spark systems.
200 GBPS CONNECTX-702 / THE NEXT THING YOU BUILD
What’s your next move?
Choose a workload to explore where Spark fits.
UP TO 200B PARAMETERS
Make room for larger models.
Explore local inference with models of up to 200 billion parameters. Capacity depends on precision, architecture, and runtime overhead.
Explore NVIDIA playbooks ↗01workload: local_inference
02engine: NVIDIA AI software stack
03memory: coherent_unified
04data_location: your_desktop
05_
03 / THINK IN MEMORY
Your model.
Your headroom.
A simple weight-memory calculator. Explore how model size and precision change the memory equation.
Illustrative only. Includes an arbitrary 16 GB reserve for software and working memory. Actual KV cache, quantization overhead, context length, and framework requirements vary. A fit here does not guarantee a supported or performant workload.
51.0 GBILLUSTRATIVE MEMORY / 128 GB
77.0 GB remaining in this simplified estimate.
