RP Tech Demos NVIDIA DGX Spark to 100+ Bangalore Developers
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- RP Tech, an NVIDIA Partner, hosted a developer demo of the NVIDIA DGX Spark on April 11, 2026, at Hyatt Centric, Bengaluru, drawing more than 100 experienced developers for a live technical walkthrough
- NVIDIA's Amit Kumar demonstrated the DGX Spark's 128 GB unified memory, where the Grace CPU and Blackwell GPU share a single memory pool via NVLink at approximately 300 Gbps, eliminating the CPU-GPU PCI bus bottleneck
- Kumar ran a live inference of the Nemotron 3 120-billion-parameter model, quantized to 86 GB via Ollama's format, with the room watching GPU utilization climb to 95% and then return to zero entirely on the tabletop device
- NemoClaw was presented as NVIDIA's enterprise version of the open-source Open Claw agent, running inside a deterministic sandbox created by Open Shell where YAML-defined policies block unauthorized actions regardless of prompt content
- DGX Spark uses LPDDR memory at 150 watts and is not pitched as a data center replacement, but two units can be linked via QSFP cable to form a 256 GB unified memory system
- Fine-tuning scope on the device supports LoRA fine-tuning up to 70 billion parameters and full fine-tuning up to approximately 13 billion parameters
Why it matters: NVIDIA is positioning the DGX Spark as a credible local alternative to cloud-only AI development — a 1.2 kg machine that runs 120B-parameter models and enforces enterprise-grade agent boundaries without data leaving the desk. For the 100+ developers in the room and the broader Indian ecosystem RP Tech is targeting, the demo reframes where serious AI prototyping and agent deployment can happen: on hardware that costs no recurring inference fees.
Ask SkimNews
