How NVIDIA engineers and researchers build with Codex
Leverage Codex with GPT‑5.5 on NVIDIA GPUs to accelerate research workflows, automate code generation, and prototype production systems.
Integrate: Run Codex on NVIDIA infrastructure for new research projects, and benchmark speed gains against existing pipelines.
Summary
NVIDIA engineers use Codex built on GPT‑5.5 to accelerate research workflows, prototype production systems, and automate code generation.
Codex runs on NVIDIA GB200 and GB300 GPUs, handling long autonomous sessions and surfacing bugs and gaps that earlier models missed.
The team reports a 10× speed improvement in end‑to‑end machine‑learning research workflows and uses Codex to evolve an internal platform from MVP to production.
Codex also built a podcast‑recording app in hours, automatically testing video and audio functionality without developer intervention.
The Codex desktop app supports SSH, allowing researchers to run large ML workloads from their laptops without manual login setup.
Codex can translate Python repositories into Rust, improving performance 20×, and acts as a research agent that writes scripts, runs experiments, and visualizes knowledge graphs.
NVIDIA plans to continue expanding Codex across engineering and research teams, pushing the boundary of what can be built in a single workflow.
Key changes
- Codex runs on NVIDIA GB200/GB300, handling long autonomous sessions and surfacing bugs and gaps.
- 10× speed improvement in end‑to‑end machine‑learning research workflows.
- Automates code generation, bug detection, refactoring, and prototype production systems.
- Built an internal podcast‑recording app in hours with automated testing.
- Supports SSH for remote ML workloads, eliminating manual login setup.
- Translates Python to Rust, improving efficiency 20×.
- Acts as a research agent that writes scripts, runs experiments, and visualizes knowledge graphs.
- Enables rapid prototyping of production systems.