Pedro Torruella

Pedro Torruella is a senior software engineer on NVIDIA’s software compilers team, specializing in GPU kernel optimization and machine-learning techniques for the GPU programming stack. He co-authored NVIDIA CompileIQ, a CUDA 13.3 optimization framework that automatically tunes compiler decisions to individual workloads.

Torruella began his career in real-time video processing, implementing algorithms in hardware and software. He subsequently managed programs, led multinational teams, and founded a startup. He holds a joint master’s degree in electrical and computer engineering from the University of Southampton and RPTU Kaiserslautern.

At OctoAI, he worked in developer relations, helping engineers deploy open-source language models, build retrieval systems, and integrate customized models into production applications. He authored the LlamaIndex OctoAI embeddings integration, which supports individual and batched embedding requests, and independently presented on production language models and open-source providers.

At AI Engineer World’s Fair 2024, he supported an OctoAI fine-tuning workshop led by company co-founder Thierry Moreau, helping troubleshoot model-asset deployment during a demonstration involving Llama 3, function calling, and sensitive-data redaction. He also co-hosted an OctoAI agentic AI and function-calling webinar.

• Workload-specific compiler auto-tuning: NVIDIA CompileIQ uses evolutionary search to explore register allocation, instruction scheduling, and loop transformations, producing kernel configurations tailored to actual workloads.
• Multi-objective performance optimization: Its search weighs execution speed against compilation time and power consumption, giving engineers practical choices for different deployment constraints.
• Production-ready model infrastructure: His OctoAI work linked embeddings, retrieval, model customization, function calling, and deployment into concrete developer workflows.