MediaTek

Solution

MediaTek built an on-prem AI factory around NVIDIA DGX SuperPOD to centralize model training, inference, operations, and developer experimentation for semiconductor R&D. Blackwell-based DGX infrastructure gives the company the scale to train 480B-parameter traditional-Chinese models within a week, process tens of billions of inference tokens each month, and host a high-demand internal API that later distills smaller models to edge and mobile use cases. NVIDIA Mission Control streamlines provisioning, deployment, health checks, and resource monitoring across the DGX cluster, while NVIDIA NeMo fine-tunes domain-specific models. For serving, MediaTek uses NVIDIA NIM with TensorRT-LLM to lift inference speed and token throughput, and NVIDIA Riva accelerates ASR and TTS in its agentic voice-control flows. The same platform also powers AI-assisted code completion and documentation generation, reducing engineering friction while keeping proprietary data on local systems.

Data flow

Engineering data and model-development workloads run on MediaTek’s on-prem DGX SuperPOD, Mission Control provisions and monitors the cluster, NeMo tunes domain models, NIM and TensorRT-LLM serve them through centralized APIs, and the outputs flow into code-completion, documentation, inference, and voice-agent workflows.