NVIDIA

How NVIDIA Engineering Teams Cut Time to Merge

6 hrs
Average time to merge, down from 24+ hrs
395K+
Pull requests reviewed across all NVIDIA repos

See how Greptile can help your team.

Industry: Semiconductors
Stage: Public (NASDAQ: NVDA)
Tech Stack: C++, Python, CUDA
Git Platform: Perforce, GitLab
NVIDIA logo

Code reviews happen very quickly because Greptile is the first code reviewer. We're spending a lot less time doing code reviews, and are much more productive in shipping code.

Todd Tanber

Senior Engineering Manager, Diagnostics, Architecture and Infrastructure, NVIDIA

Todd Tanber


Q&A

Todd Tanber, Senior Engineering Manager: Greptile stood out for a couple of reasons. Primarily, Greptile handles large codebases really well.

We also liked the cross-file graph index context. Most code review tools just look at the diff — Greptile reads across the entire codebase. That's essential for NVIDIA at the scale in which we operate. It also identifies bugs that we're going to see out in the future — it recognizes those up front.

Todd Tanber, Senior Engineering Manager: Greptile offered support for Perforce, which is a major source code management (SCM) tool at NVIDIA. Working deeply between Greptile and NVIDIA developers, Greptile delivered 35 custom features for our Perforce instance, as well as 55 custom features for GitLab instance. That was a great win for NVIDIA.

When we look further, Greptile operates independently of the code generation layer. So the tools that are writing the software are not the tools that are reviewing the software. That's yet another positive element.

Praveen Dorairaj, Senior Software Tools Infrastructure Architect, Hardware Infrastructure: Before Greptile I was drowning in code reviews — six people, an endless stream of reviews that come out. A large portion of my time was spent reviewing everyone's code. And because of this, I would miss logic errors here and there. You can't catch everything at that level of review.

I would frequently see that our coding patterns were not adhered to. I would see a lot of basic logic errors. This would cause reviews to go through many more iterations through me, which just increases the amount of time it takes to finish the review of code.

But now with Greptile, I can be sure that at least when I get to the code, a couple of iterations have been passed. The code should be more stable, and at least our coding standards should be adhered to.

Todd Tanber, Senior Engineering Manager: Greptile allows NVIDIA-specific coding standards — almost like a style convention, if you will — and it gets everybody working to the same sort of expectations. That further helps as we bring on new team members. It's less training required. Everyone's working already, those expectations are already set, and the team members get to work within the same guidelines everyone else is operating to.

Peiyuan Qi, Hardware Infrastructure Software Engineer: Before using Greptile, our team's average code review time to merge a merge request was more than 24 hours. Nowadays I check my dashboard — the average merge request close time is as low as six hours. Since we adopted Greptile in April, we've already merged more than 1,300 merge requests for this product. Greptile makes our team more productive and helps us to move fast.

Praveen Dorairaj, Senior Software Tools Infrastructure Architect, Hardware Infrastructure: I used to be skeptical of AI code reviews at first — like, what can this model really know about my codebase? After using it for a while, I would say Greptile is very adaptable, very configurable, so you can go in and tweak what you feel would help you the most. Especially if you're drowning in code reviews and you feel like that's taking up most of your time.

Todd Tanber, Senior Engineering Manager: Greptile has always gone the extra mile to deliver to the needs of NVIDIA. The feedback we get from developers is strongly positive in terms of the Greptile reviews.


[ CODEBASE SCALE ]

What it took to review code at NVIDIA

NVIDIA chose Greptile to review across Perforce and GitLab at scale. One team's time to merge went from more than 24 hours to six.

NVIDIA writes C++, Python, and CUDA across Perforce and GitLab. Greptile built 35 custom features for the Perforce instance and 55 for GitLab.

Most review tools read the diff. Greptile indexes across files, so a change gets reviewed against the callers and contracts it doesn't touch. It runs independently of the code generation layer — the tools writing the software at NVIDIA are not the tools reviewing it.

One team's average time to merge went from more than 24 hours to six. Since April they have merged more than 1,300 merge requests. That is one team on one product. Across NVIDIA's self-hosted deployment, Greptile has reviewed roughly 395,000 pull requests over roughly 880,000 review runs, and posted roughly 930,000 review comments.



[ KEEP READING ]

Learn more about Greptile


Greptile consistently catches more bugs in large repos