A cookbook for deploying GenAI across Intel, AMD, and NVIDIA

What it does

GenAIExamples is a catalog of ready-made generative AI applications—chatbots with RAG, document summarizers, code generators, and visual Q&A systems—packaged as microservice-based reference architectures. Each example is wired together from the companion GenAIComps library and ships with deployment manifests for Docker Compose and Kubernetes. The project treats hardware as a configuration axis: the same pipeline can target Intel Xeon, Intel Gaudi, AMD EPYC, AMD Instinct, or NVIDIA GPUs.

The interesting bit

The real work here is not the models but the integration matrix. Maintaining parity across Helm charts, GMC manifests, and Docker Compose files for Xeon, Gaudi, AMD EPYC, and AMD Instinct is the kind of tedious, high-leverage plumbing that most teams end up rebuilding privately. OPEA open-sources that grind.

Key highlights

• Covers use cases: ChatQnA (RAG), DocSum, CodeGen, Text2Image, VisualQnA, and instruction tuning.
• Deployment targets span bare-metal Python startup, Docker Compose, and Kubernetes via Helm or GMC.
• Hardware support explicitly includes Intel Gaudi and Xeon, AMD EPYC and Instinct MI300X, and the project notes compatibility with NVIDIA GPUs.
• Evaluation is built into the ecosystem via the separate GenAIEval project, which tracks throughput, latency, and accuracy.
• The README is essentially a large compatibility matrix linking to per-example, per-platform setup guides.

Caveats

• The repository is mostly glue and configuration: the actual microservices live in GenAIComps, while deployment automation and evaluation live in GenAIInfra and GenAIEval.
• Some combinations are unsupported; for example, SearchQnA lacks ROCm deployment instructions.
• The README offers hardware sizing guidance, but actual performance numbers or benchmarks are not included in the repository itself.

Verdict

Grab this if you are an infrastructure or platform engineer deciding how to standardize GenAI deployments across a mixed CPU/GPU fleet. Skip it if you are looking for a novel model architecture or a single, self-contained application.