Innovation Through Research

DXG Spark helped to predict protein structure changes for ~6,000 epilepsy-related mutations locally using Boltz‑2. It serves as a rapid iteration “workbench” for screening small proteins and an inhibitory receptor, validating runtimes on a single GPU before scaling insights to the Kempner AI cluster (“launch 100 GPUs to test 100 proteins in a few days”). This desktop-to-cluster flow compresses research cycles by letting biologists test ideas, write code, and refine models at their desks, then expand to larger runs as needed.

DGX Spark is the engine behind NVIDIA’s Sionna Research Kit: a real-time, open 5G/6G lab-in-a-box. On a single system, you can run an OAI base station, emulate RF with hardware ray tracing, accelerate PHY (LDPC on GPU) and neural receivers (TensorRT), and monitor with RIC xApps—all on unified memory, no fixed-function ASICs. Scale the same code to DGX Cloud for continent-wide radio maps (35T rays on 96 L40S GPUs in <5 min).

LATTICEPT was able to run in DGX Spark incredible multiphase computational fluid dynamics simulations with DEM coupling. LATTICEPT does all sorts of FEA and CFD consulting in oil, gas, pharmaceutical, mining, chemicals and lots more

With NVIDIA SPARK DGX, WebRobot enters a new stage of agentic computing — one where data flows, inference, and fine-tuning can all happen locally, without worrying about token consumption or latency. This is not just hardware — it’s a paradigm shift toward the democratization of AI.

DGX Spark dramatically accelerate the development of Mobiltech's digital twin platform, Replica City. With seamless compatibility with NVIDIA Omniverse and OpenUSD, capable to process massive 3D spatial data rapidly and implement highly advanced AI simulation environments.

For years, foundation models have revolutionized text (GPT) and images (DALL-E), but enterprise data has been left behind. SAP customer's business-critical insights live in relational databases—customers linked to orders, products tied to inventory, users connected to events across time. Traditional ML approaches require lots of feature engineering, taking weeks to deliver results.