NVIDIA DGX Spark: GB10 Connected Intelligence

A voice assistant feels natural only when it answers as fast as a person talks. Several models have to work in turn: one hears, one thinks, one speaks and one animates a face, and on a single machine they compete for the same memory bandwidth. This lab shows how two NVIDIA DGX Spark systems split that work so the conversation stays real time, with no cloud in the loop.

You meet Sky, a real-time AI assistant, and its host persona Andrew on a video call. Spark 1 runs the conversation: speech recognition, NVIDIA Nemotron 3.5 Lightning writing each answer, and a live video face whose lips follow the voice. Spark 2 speaks the answer in Andrew's cloned voice and holds the knowledge Sky searches, with room for a deep model beside them. The two are joined by a direct 200 GbE link.

You interrupt Andrew mid-sentence, watch each turn cross both Sparks in Under the Hood and System Overview, and see what the second Spark buys (isolation and headroom) and what it does not (pooled memory or graphics). Optional modules show how a second persona, Samantha, was built from two video clips and a voice reference, and how Sky runs a simulated building through Home Assistant, with a confirmation before anything consequential.

Everything runs in the browser through the lab gateway: nothing to install and no VPN.

• Hold a real-time video call with an AI assistant that runs entirely on two DGX Spark systems, with no cloud in the loop
• Interrupt the assistant mid-answer and see how turn-taking, barge-in and session history work
• Trace one spoken turn across both Sparks: speech recognition, the language model and the live face on Spark 1, the voice and the knowledge search on Spark 2
• Read live service, memory and timing data in System Overview and Under the Hood
• Explain what a second Spark buys (isolation, headroom, independent servicing) and what it does not (pooled memory, graphics)
• Optional: see how a synthetic persona is built from two video clips and a voice reference, with consent and provenance, and how personas stay scoped
• Optional: have the assistant read and control a simulated building, with confirmation before consequential actions

Hardware

• 2x NVIDIA DGX Spark, each with one NVIDIA GB10 Grace Blackwell Superchip and 128 GB of unified memory
• Two direct 200 GbE NVIDIA ConnectX-7 links between the Sparks

Software on Spark 1 (conversation)

• NVIDIA Nemotron 3.5 Lightning 30B-A3B (NVFP4) on vLLM, with speculative decoding
• Whisper large-v3-turbo (speech to text)
• NVIDIA Audio2Face-3D and Ditto (lip sync)
• LivePortrait on TensorRT with NVENC video encoding (the live video face)
• Home Assistant with simulated buildings
• SkyOS, the assistant's web interface

Software on Spark 2 (voice and heavy models)

• CosyVoice 3 (zero-shot voice cloning text to speech)
• Knowledge service with hybrid vector and keyword search
• A second Nemotron 3.5 Lightning as a deep model, behind a voice-first gate
• Lab Buddy (the guide's setup, validate and solve checks)

You need

• A modern browser (Chrome, Edge or Safari) with a microphone, ideally a headset. Nothing to install.