---
格式版本: 2
标题: "Practical Context Engineering: Eliminate Bugs With High-Signal AI Code Reviews S81612 | GTC San Jose 2026"
原文链接: "https://www.nvidia.com/gtc/session-catalog/sessions/gtc26-s81612/"
发布日期: "2026-09-07"
发布时间校准状态: "found"
发布时间需复核: "否"
发布时间来源: "rule:local:strict_original_body"
发布时间证据: "Published Time: Mon, 07 Sep 2026 15:22:56 GMT"
发布时间校准原因: "规则确认唯一严格发布时间，来源 local:strict_original_body"
发布时间校准置信度: "high"
发布时间候选数量: 4
发布时间严格候选数量: 1
发布时间原页读取状态: "source template page reused from URL open"
发布时间未找到原因: ""
发布时间校准时间: "2026-09-07T23:23:02+08:00"
发布时间仲裁状态: "skipped"
发布时间仲裁尝试次数: 0
发布时间仲裁耗时毫秒: 0
发现时间: "2026-09-07T23:16:22+08:00"
入库时间: "2026-09-07T15:26:43.786Z"
来源平台: "固定入口"
搜索渠道: "fixed_url"
搜索词: "https://www.nvidia.com/gtc/"
匹配关键词:
- "performance"
- "bandwidth"
- "throughput"
- "AI"
相关厂家:
- "NVIDIA"
相关专家:
[]
内容类型: "网页"
抓取工具: "Jina Reader"
清洗工具: "Jina Reader Markdown + Defuddle/Readability 正文提取"
原始附件:
[]
AI优质: "否"
AI打分: 26
AI分档: "非优质"
AI质检状态: "不通过"
AI打分理由: "正文为GTC 2026 CodeRabbit演讲转录，主线是AI代码审查与上下文工程，属于软件开发工具/最佳实践，不讨论超节点、AI Rack或机架级基础设施，无新机架产品、标准、部署或技术路线。来源为NVIDIA官方会话页，完整可追溯，但主题不相关。命中教程与运维选型及应用/模型效率否决项，最高54分，判false。"
AI质检模型: "zj-deepseek-v4-flash"
AI质检时间: "2026-09-07T23:27:01+08:00"
AI主题相关性: 2
AI来源权威性: 10
AI新颖性: 3
AI技术细节: 2
AI商业部署信号: 2
AI完整性: 7
AI评分提示词版本: "v17-精简生产版"
AI评分提示词SHA256: "48fb9777f386026761b4873eaff30807694fb11e9b352d7c69bf2dfde750cc7d"
AI评分知识库版本: "knowledge_base_v1-20260819+runtime.100"
AI评分知识库SHA256: "7fe4568b9de9fc322c105ece08d18efab5b9b86f27dea6eb043b9a4d273bde60"
AI评分知识库检索词: "[\"NVIDIA\",\"Nvidia GTC大会\",\"https://www.nvidia.com/gtc/\",\"RAS\",\"NPU\",\"S81612\",\"GTC\",\"URL\",\"GMT\",\"GTCs\",\"EN\",\"CI\"]"
AI评分知识库命中: "[{\"id\":\"runtime-47bb908e0d4e4e5fe3587a31\",\"title\":\"View On-Demand\",\"sourceType\":\"ai_excellent_article\",\"time\":\"\",\"matchedTerms\":[\"NVIDIA\",\"https://www.nvidia.com/gtc/\",\"RAS\",\"GTC\",\"URL\",\"GMT\",\"GTCs\",\"EN\",\"CI\"],\"rank\":-16.827927561563087},{\"id\":\"runtime-a63fdca649ea6d74b79483ba\",\"title\":\"View On-Demand\",\"sourceType\":\"ai_excellent_article\",\"time\":\"\",\"matchedTerms\":[\"NVIDIA\",\"https://www.nvidia.com/gtc/\",\"RAS\",\"GTC\",\"URL\",\"GMT\",\"GTCs\",\"EN\",\"CI\"],\"rank\":-16.65577781893091},{\"id\":\"runtime-fa6eb5eb4c45c403e767993d\",\"title\":\"View On-Demand\",\"sourceType\":\"ai_excellent_article\",\"time\":\"\",\"matchedTerms\":[\"NVIDIA\",\"https://www.nvidia.com/gtc/\",\"RAS\",\"GTC\",\"URL\",\"GMT\",\"GTCs\",\"EN\",\"CI\"],\"rank\":-15.575342323393267},{\"id\":\"july-correct-0001\",\"title\":\"全球首颗2nm GPU来了！苏姿丰甩出“最强AI机架”，CPU性能干翻英伟达 - 智东西\",\"sourceType\":\"labeled_article\",\"time\":\"2026-07\",\"matchedTerms\":[\"RAS\",\"NPU\",\"EN\",\"CI\"],\"rank\":-11.428416651191942},{\"id\":\"july-correct-0088\",\"title\":\"StrataCL: Fabric-Native Communication Library for Production Supernodes - 智源社区论文\",\"sourceType\":\"labeled_article\",\"time\":\"2026-07\",\"matchedTerms\":[\"NVIDIA\",\"NPU\",\"URL\",\"EN\",\"CI\"],\"rank\":-9.838172516934957}]"
采集批次: "2026年9月7日23点15分57秒"
采集批次ID: "20260907-231557-722"
去重键: "https://www.nvidia.com/gtc/session-catalog/sessions/gtc26-s81612"
---
Title: Practical Context Engineering: Eliminate Bugs With High-Signal AI Code Reviews S81612 | GTC San Jose 2026
URL Source: https://www.nvidia.com/gtc/session-catalog/sessions/gtc26-s81612/
Published Time: Mon, 07 Sep 2026 15:22:56 GMT
Markdown Content:
Visit your regional NVIDIA website for local content, pricing, and where to buy partners specific to your country.
[Continue](https://www.nvidia.com/)
Upcoming: Explore other GTCs: [**GTC Berlin** October 20–22](https://www.nvidia.com/en-eu/gtc/) |
[**GTC Washington, D.C.** November 30-December 3](https://www.nvidia.com/gtc/dc/)
| [**GTC 2027** March 15–18](https://www.nvidia.com/gtc/)
* [](https://www.nvidia.com/gtc/session-catalog/sessions/gtc26-s81612/#)
* [](https://www.nvidia.com/gtc/session-catalog/sessions/gtc26-s81612/#)
* [](https://www.nvidia.com/en-us/account/)
* [Log In](https://www.nvidia.com/gtc/session-catalog/sessions/gtc26-s81612/#)[LogOut](https://www.nvidia.com/gtc/session-catalog/sessions/gtc26-s81612/#)
* [EN](https://www.nvidia.com/gtc/session-catalog/sessions/gtc26-s81612/#)
* [EN](https://www.nvidia.com/en-us/on-demand/)
* [简中](https://www.nvidia.cn/on-demand/)
* [日本語](https://www.nvidia.com/ja-jp/on-demand/)
* [한국어](https://www.nvidia.com/ko-kr/on-demand/)
* [繁中](https://www.nvidia.com/zh-tw/on-demand/)
PLATFORMS
other links
[](https://www.nvidia.com/gtc/)
Keynote
* [Keynote](https://www.nvidia.com/gtc/keynote/)
* [_GTC Live_ Pregame](https://www.nvidia.com/gtc/pregame/)
Explore
* [Conference Topics](https://www.nvidia.com/gtc/conference-topics/)
* [Poster Gallery](https://www.nvidia.com/gtc/posters/)
* [Speakers](https://www.nvidia.com/gtc/speakers/)
* [Startups & VCs](https://www.nvidia.com/gtc/startups/)
* [Workshops, Training Labs & Certification](https://www.nvidia.com/gtc/training/)
[Sponsors & Exhibitors](https://www.nvidia.com/gtc/sponsors/)
More
* [Code of Conduct](https://www.nvidia.com/gtc/code-of-conduct/)
* [Contact Us](https://www.nvidia.com/gtc/contact/)
* [FAQ](https://www.nvidia.com/gtc/faq/)
* [Privacy Policy](https://www.nvidia.com/en-us/about-nvidia/privacy-policy/)
* [See All GTC Events](https://www.nvidia.com/gtc/events/)
* Keynote
* [Keynote](https://www.nvidia.com/gtc/keynote/)
* [_GTC Live_ Pregame](https://www.nvidia.com/gtc/pregame/)
* Explore
* [Conference Topics](https://www.nvidia.com/gtc/conference-topics/)
* [Poster Gallery](https://www.nvidia.com/gtc/posters/)
* [Speakers](https://www.nvidia.com/gtc/speakers/)
* [Startups & VCs](https://www.nvidia.com/gtc/startups/)
* [Workshops, Training Labs & Certification](https://www.nvidia.com/gtc/training/)
* [Sponsors & Exhibitors](https://www.nvidia.com/gtc/sponsors/)
* More
* [Code of Conduct](https://www.nvidia.com/gtc/code-of-conduct/)
* [Contact Us](https://www.nvidia.com/gtc/contact/)
* [FAQ](https://www.nvidia.com/gtc/faq/)
* [Privacy Policy](https://www.nvidia.com/en-us/about-nvidia/privacy-policy/)
* [See All GTC Events](https://www.nvidia.com/gtc/events/)
* [](https://www.nvidia.com/gtc/session-catalog/sessions/gtc26-s81612/#)
* [EN](https://www.nvidia.com/gtc/session-catalog/sessions/gtc26-s81612/#)
* [EN](https://www.nvidia.com/en-us/on-demand/)
* [简中](https://www.nvidia.cn/on-demand/)
* [日本語](https://www.nvidia.com/ja-jp/on-demand/)
* [한국어](https://www.nvidia.com/ko-kr/on-demand/)
* [繁中](https://www.nvidia.com/zh-tw/on-demand/)
[Watch On Demand](https://www.nvidia.com/en-us/on-demand/search/?facet.event_name[]=GTC%20San%20Jose&facet.event_year[]=2026&facet.mimetype[]=event%20session&headerText=All%20Sessions&layout=list&page=1&q=-&sort=relevance&sortDir=desc>cnavinherit=true)[](https://www.nvidia.com/gtc/session-catalog/sessions/gtc26-s81612/# "Menu")
[Watch On Demand](https://www.nvidia.com/en-us/on-demand/search/?facet.event_name[]=GTC%20San%20Jose&facet.event_year[]=2026&facet.mimetype[]=event%20session&headerText=All%20Sessions&layout=list&page=1&q=-&sort=relevance&sortDir=desc>cnavinherit=true)[](https://www.nvidia.com/gtc/session-catalog/sessions/gtc26-s81612/# "Menu")
* [Keynote](https://www.nvidia.com/gtc/session-catalog/sessions/gtc26-s81612/#)
* [Keynote](https://www.nvidia.com/gtc/session-catalog/sessions/gtc26-s81612/#)
* [Keynote](https://www.nvidia.com/gtc/keynote/)
* [_GTC Live_ Pregame](https://www.nvidia.com/gtc/pregame/)
* [Explore](https://www.nvidia.com/gtc/session-catalog/sessions/gtc26-s81612/#)
* [Explore](https://www.nvidia.com/gtc/session-catalog/sessions/gtc26-s81612/#)
* [Conference Topics](https://www.nvidia.com/gtc/conference-topics/)
* [Poster Gallery](https://www.nvidia.com/gtc/posters/)
* [Speakers](https://www.nvidia.com/gtc/speakers/)
* [Startups & VCs](https://www.nvidia.com/gtc/startups/)
* [Workshops, Training Labs & Certification](https://www.nvidia.com/gtc/training/)
* [Sponsors & Exhibitors](https://www.nvidia.com/gtc/sponsors/)
* [More](https://www.nvidia.com/gtc/session-catalog/sessions/gtc26-s81612/#)
* [More](https://www.nvidia.com/gtc/session-catalog/sessions/gtc26-s81612/#)
* [Code of Conduct](https://www.nvidia.com/gtc/code-of-conduct/)
* [Contact Us](https://www.nvidia.com/gtc/contact/)
* [FAQ](https://www.nvidia.com/gtc/faq/)
* [Privacy Policy](https://www.nvidia.com/en-us/about-nvidia/privacy-policy/)
* [See All GTC Events](https://www.nvidia.com/gtc/events/)
[Video 1](blob:https://www.nvidia.com/d402a814-ca23-4524-8afd-19027f55a925)
Loading
Play
00:00
Play
Seek 10 seconds backwards
Seek 10 seconds forward
00:00 / 00:00
Mute
Press TAB to open volume control Use the arrows to control the volume
Enable Captions
Settings
Play on TV
Turn on Picture in picture
Show Full screen
Transcript Powered by AI
X
X
00:06
Thank you very much.
00:10
Yeah, so I'd say, yeah, coding is definitely one
00:12
of the main use cases we're seeing, obviously, proliferate
00:15
at this particular point.
00:15
I don't know if anyone has not touched one of the systems
00:19
at this point to generate code.
00:21
And now everybody, yeah, everybody across the board is generating
00:23
code on some level.
00:24
And so, obviously, as that happens, code reviews become
00:28
even more important, right?
00:31
We now live, right, in this whole new ecosystem.
00:35
AI is driving just about everything.
00:36
We have groups of people who are saying, I haven't written the
00:39
line of code in the last six months or since the beginning of the year.
00:44
And so how do we adapt?
00:45
How do we change the way that we work in order to accommodate
00:49
this new way of working?
00:52
And so just looking back, right?
00:55
Handwritten code, I'm a pre-AI, I have a bunch of CI CD pipelines,
00:59
tests, linters, a whole bunch of things just to try and
01:02
make sure, again, that what we're writing is good, is solid,
01:05
when it goes into production, it's not going to break everything, and
01:08
then we move into this AI-assisted coding, right, so we have Copilot,
01:11
ChatGPT, Cursor, and others, Codecs, and going into the future,
01:20
And AI-driven, where we're just using agents, right?
01:23
In the background, I say what I want or what I'm
01:25
trying to accomplish.
01:26
Maybe I provided a lot of business context of what's going on.
01:30
And then I walk away and I let these set of agents go
01:32
off and I let them code everything.
01:34
How do I know what comes out of it?
01:36
How do I validate what comes out of it?
01:38
How am I sure that it didn't build a lot more or a lot
01:40
less than what I wanted?
01:41
How am I sure that I was specific enough?
01:45
In my ask, so that I actually knew what to build rather than
01:47
making assumptions that was against what I was trying to go for.
01:51
And so we have this new, this new way, right?
01:55
This new ecosystem, code generation, super fast,
02:00
code review, super slow, right?
02:03
We're generating a lot of code.
02:05
If I have to read that 15,000 line PR that you just opened with your
02:08
background agents, I'm not going to get through it and I'm not going
02:12
to be able to do a good job, right?
02:14
And so this old way of doing things, we have
02:18
a new bottleneck, right?
02:19
We have a new bottleneck in the review and the validation
02:22
portion of the workflow.
02:27
Just to get a little bit more concrete, a billion lines
02:33
of code generated per day, this is a little bit old even from Cursor,
02:37
30% increase in the number of pull requests of people again shipping
02:42
features at a higher velocity.
02:46
And then we have a big increase in the number
02:47
of bugs within those PRs.
02:49
So we're actually compounding the problem.
02:51
Not only do we have slower bottlenecking reviews, but we
02:54
actually have more things to find.
02:57
And so that is a massive scalability challenge when
02:59
it comes to how do we make sure, again, that our production
03:02
systems aren't breaking.
03:07
Previously, we had 15% to 25% of a developer time going
03:10
into code review.
03:11
And this is some of your more expensive resources, even.
03:13
You have your senior and architect-level engineers
03:16
getting in there, trying to make sure that the quality bar is high.
03:21
But it's 100% 100 times more expensive to fix it if it
03:23
gets to production.
03:24
So you want to spend that effort.
03:26
You want to spend that resource, because you don't want those
03:29
things to happen in production.
03:30
And in some cases, enterprises, you're talking about up to$
03:33
5 million per hour of downtime.
03:35
So this is a really big problem that needs to be solved, right?
03:38
That's the cost of bugs.
03:42
And so, should we just let Cloud Code or Cursor review
03:47
their own PRs, right?
03:48
So, they've already generated 41% more bugs in their code.
03:52
So, should we really trust them to find those bugs after
03:54
they just generated them?
03:57
They tend to be optimized for speed, right?
03:58
If I'm sitting there at my CLI and I'm using Cloud Code
04:01
or I'm using Cursor, they're meant to be interactive.
04:04
They're meant to engage with a human being.
04:05
And so they don't necessarily take the time required to actually
04:09
find those hard to find bugs.
04:13
They're not focused on all the features that you would need
04:16
as part of a code review because it's not their main product.
04:19
And not everybody uses the same tool.
04:22
So the quality and the differences in the review is going to
04:25
shift depending on the developer.
04:26
And so as an organization, if we're trying to set some sort
04:29
of quality standard, we sort of lose the ability to do that, right?
04:34
We want to have some sort of consistency in the bar
04:38
that we're trying to set.
04:41
And so we have this new, what we consider the new AI DevTools stack.
04:44
You have your code generation systems, and they might be
04:47
agentic, they might be fully in the background.
04:49
You might be having them sort of asynchronous.
04:53
And then we have your central...
04:56
Repository, GitHub, GitLab, Bitbucket, Azure DevOps,
05:00
these sorts of things.
05:01
And you need some sort of AI code review system in order
05:04
to help manage the amount of code that's coming in and
05:08
the proliferation of bugs that are being inserted into your code
05:13
base by these new AI-generated PRs.
05:18
So CodeRabbit obviously sits there, and that's what I'm
05:20
going to talk about today.
05:21
So we're used by a lot of different companies across
05:24
the board, including NVIDIA.
05:28
We're a Series B company.
05:29
We've raised over $88 million in total.
05:31
We have over now 250,000 monthly active users, 75 million plus
05:36
defects found, and we've reviewed over 3 million pull requests.
05:44
There are three ways that you can use CodeRabbit inside your system.
05:48
If you're a developer and you're on your laptop and
05:50
you want to use it in the IDE or in a CLI tool, maybe even
05:54
feed it directly into Cloud Code, there's a plug-in that you can
05:58
just hook CodeRabbit directly up.
06:00
And ultimately, we also live in the central layer, which
06:02
is, I think, one of the most important places for something
06:04
like AI code review to do is to live in that central layer so that
06:08
you can make sure that everything that goes into production
06:11
has been evaluated by some tool that's looking for these bugs.
06:17
And so that works across all those different platforms.
06:21
Now, the way that we do it, okay, so how do you
06:24
deal with code review?
06:26
Getting into some more interesting points of the conversation.
06:30
Every task, right, that an AI has to do, and you'll hear
06:33
this over and over again over the past, I'm sure, six months
06:35
or so, is that context is king.
06:38
What would I need to do to be able to review your PR?
06:42
I need information.
06:43
If I've been at the company a while, I have a lot of accumulated
06:45
information, but even still, if I look at code that I wrote
06:48
six months ago, I might not even recognize what I did.
06:51
So I need to go and need to understand it.
06:53
I need to go and I need to look around, right?
06:55
I might need to read the issue.
06:56
How did this come about?
06:57
What are we building and why?
06:58
I need to look at the surrounding code.
07:01
If you changed a function signature in some way and it's used
07:04
elsewhere, I need to understand, did you break something?
07:07
If I'm a consumer of that, did you change the core functionality
07:11
that would break some sort of contract between the consumer
07:14
and now the producer?
07:16
And so there's a lot of context enrichment.
07:18
The majority, the vast majority of the work done is in the
07:22
context enrichment side of things.
07:24
And we can use a lot of domain expertise in order to make
07:27
sure that that context enrichment is happening in a way that
07:30
would effectively make me able to review that PR.
07:34
So we have to clone the repository.
07:37
We have to do this thing and we do code graph analysis.
07:39
How your code and your PR is connected up to the other
07:42
aspects of the code base.
07:45
Issues, so I was talking about earlier, why?
07:48
Why did you build this?
07:49
What was the purpose of it? What were you trying to accomplish?
07:52
You can connect MCP systems, so for example, your architecture
07:55
documentation, your coding guidelines, things that your
07:58
standards that you might have, I should evaluate against that, make
08:01
sure that everything is adhered to.
08:03
Maybe you have a specific way that you do encryption,
08:05
making sure that that's adhered to as well, or security posture.
08:10
And so then you have other things like you're using
08:12
coding agents, right?
08:13
You're using Claude, you're using Cursor.
08:15
You have these MD files.
08:16
You spent a lot of time putting a lot of information for how these
08:19
systems should work into those.
08:21
So we need to take those. We need to make sure that the code
08:24
being generated actually adheres to those guidelines that you set.
08:29
And then we do other things.
08:30
There's personalization inside code review.
08:33
One company might never want variables to be public in a class.
08:38
Another company doesn't care.
08:40
So which one, do I show the error or do I not?
08:43
Ultimately, you wanna make sure our product is geared towards finding
08:46
as many bugs that exist as possible and then letting you decide.
08:52
What are the ones that you want to see on a regular basis?
08:54
And so what happens is you chat with CodeRabbit, you
08:57
tell it, oh, we don't do this in my company, this is something I don't
09:00
care about, and it learns over time so that the reviews become
09:04
more and more personalized to whatever, however you work, right?
09:10
And so there's a lot of other stuff that goes in there.
09:11
Web queries is a great one.
09:12
So libraries change all the time.
09:14
Sometimes they're breaking changes.
09:16
Models obviously are out of date when it comes to that information.
09:19
And so we need to go out and we need to look at the
09:21
documentation to make sure that you're adhering to
09:23
whatever version that you're using.
09:27
And then we go through this whole loop.
09:28
Yeah, it's got a review and then we have a verification system to make
09:31
sure that what came out is grounded in truth and fact and there's
09:34
no false positives coming out.
09:38
And then finally, we get to output the comment.
09:41
I would say 80 to 90% of token usage is in the actual context
09:44
enrichment itself.
09:46
80% of our actual usage.
09:49
The review agent itself is, yeah, it's a very heavy reasoning
09:51
task, but all that heavy lifting of finding what it needs to
09:55
actually do the job has already been done by that point.
10:03
I just want to talk a little bit about the evolution of
10:05
prompt engineering and what goes into some of this.
10:06
So obviously, to start off, prompt engineering, you know,
10:10
it asks you to do something, it fails, or it doesn't give
10:12
you exactly what you want, and so you give it some examples, right?
10:15
So this is the idea.
10:16
I'm going to do a few-shot prompting.
10:19
Little tiny changes. Ultimately, it's a very static prompt
10:21
at the end of the day.
10:22
It solves some tasks fairly well, but it doesn't really
10:24
adapt to dynamic contexts.
10:28
So that's why context engineering comes from.
10:30
This is designing the input environment so that the model has
10:34
what it needs to reason accurately.
10:36
So it's combining prompt engineering with dynamic
10:41
context gathering.
10:42
Something that's gonna sort of provide all the
10:44
information necessary.
10:47
And so that's how we fit into all this.
10:49
And I wanted to call out sort of, because this is NVIDIA and
10:53
a Nematron event, is that Nematron actually can sit in our self-hosted
10:57
environments in this middle layer that's used a lot to drive
11:01
the actual information that the reasoning system ends up consuming.
11:07
And so this is where you have your clients, your CLI, your extension,
11:11
your GitHub repository coming in, all that context gathering
11:14
takes place, and we're using Nemotron to be able to synthesize
11:18
that information and then handing off to Claude or GPT, in which
11:23
some cases Opus, some cases Sonnet, some cases GPT 5.4 or 5.3 Codex,
11:29
in order to do the actual heavy lifting of the reasoning tasks.
11:37
To get a little bit concrete on some of the things that we do
11:39
to give you an idea of the breadth,
11:42
And how you have to think about your problem space and
11:45
use your domain expertise to try and figure out what
11:48
is the context that's necessary.
11:50
And luckily, CodeRabbit, we're in an engineering space.
11:55
We have engineers driving the product.
11:57
Engineers are familiar with the code review process.
12:01
And as we view where it does well and doesn't do well, we
12:05
kind of get ideas behind, oh, what information did we maybe miss that
12:09
we could put in there that would help the LLM figure out the bug?
12:13
And so we have a whole knowledge-based system.
12:15
So some of it's personalization.
12:17
You can set up your own path, what we call path instructions.
12:20
You say TypeScript files, I want them to be reviewed
12:23
like this because I have something I do unique.
12:26
We index past PRs so that if you're fixing something
12:29
and it's related to a past PR and something went wrong over there
12:32
or something is unique, we can make sure that that same problem doesn't
12:36
occur here or that the information is correctly pulled out.
12:41
Your coding guidelines, like I referred to, the learnings
12:43
that you're going through, we have an index of the code
12:46
base and synthesized form so that, again, we can find
12:49
if you're doing authentication work, where else is authentication
12:51
happening, what is the way that you do it within your organization.
12:56
And we also now have something that's multi-repo, which is
12:58
that we actually pull information from, so say you're doing
13:01
front-end work, we can pull information from your back-end
13:04
repository and make sure that your front-end is not violating
13:07
some usage of the back-end systems.
13:09
Or if you update the back-end, that now you're not breaking
13:12
something on the front-end.
13:17
So I was talking about this earlier.
13:18
This is an example of it.
13:21
We want to get rid of star imports.
13:23
This is a comment made by someone on a PR in an open source.
13:26
And then CodeRabbit says, OK, understood.
13:29
This is a past preference that they had that you would
13:32
collapse multiple imports and you just use star instead.
13:35
So we gave that comment.
13:36
They came back saying, oh, we don't do that anymore.
13:39
And so now we added a new learning and actually removed the old ones.
13:42
So this new learning says.
13:44
It now prefers to avoid wildcard and actually instead use explicit
13:48
imports, so that in the future we don't make that same comment.
13:51
And if someone does use a star import in the future,
13:53
we pull that in and it says, don't do that, right?
13:57
The idea is that to figure out how your company works.
14:02
This is a very, very heavily used feature.
14:09
All right.
14:12
We added MCP. MCP systems is used all over the place,
14:15
but the main problem with them, if you've ever used them in any
14:17
kind of agentic context, is that the context window gets bombarded
14:22
by information that may or may not be relevant, because something
14:26
being somewhat semantically similar does not mean that it is useful
14:30
in the context that you're in.
14:32
And so when we bring in information from your MCP gateway, we're
14:36
searching for data on things like your architecture, your security
14:40
posture, or anything that might be like a PRD document we might be
14:43
searching for related to this PR.
14:46
We have to go through that then and we have to say, okay,
14:48
is there information in here that's actually relevant to
14:51
the task that I'm doing right now?
14:53
And we have to extract that information so that we're
14:55
not overloading the main context window with extraneous stuff.
14:59
Because as everybody knows at this point in time, the more
15:02
information you put into an LLM seems like a good idea at first.
15:06
But then it starts forgetting things, right?
15:08
Things in the middle start to get lost.
15:10
If I have too many instruction sets, too many things that
15:13
it has to keep track of, by the time it gets to the end, some
15:16
of that information is just lost.
15:18
So this is actually an optimization problem on top
15:20
of just finding the information.
15:22
We now need to optimize for the task at hand.
15:25
And this is why context engineering is difficult.
15:28
I can't just throw everything in there.
15:32
We have linters, ultimately, they come through and
15:35
they do simple stuff.
15:37
Oh, there's a Bandit B303 error, right?
15:41
And sometimes there's a lot of false positives in here, but
15:43
if we feed them to the LLM and use a reasoning system on top of it, we
15:46
can get rid of a lot of the false positives, but we can also explain
15:50
in the cases where it's true, why it's true, so that, again,
15:53
people can make fixes very easily.
15:56
We can actually recommend patches to fix those issues.
16:00
But this is one of the biggest ones, and this is where it doesn't
16:05
just help to put in the diff.
16:08
So the barrier to entry to code review is very small, right?
16:12
But thinking about all the different ways in which something
16:15
can break requires you to start thinking about outside the PR and
16:19
thinking how things are connected, and figuring out that information
16:22
of how things are connected, potentially multiple orders deep,
16:24
depending on the type of issue.
16:27
Like a concurrency issue might come from many levels
16:30
in order to hit your PR.
16:33
And so this is where outside diff impact slicing comes, right?
16:36
So if I look at this example,
16:38
I changed, say, for example, the process refund and
16:41
added an exception.
16:43
The handler doesn't know that, and it wasn't updated, and
16:47
so now the handler has a bug.
16:48
Because when that exception gets raised, it's just going
16:51
to break the system.
16:53
And so this is where, if you fed that to an LLM, it'd be
16:55
like, you should probably handle this error in here.
17:03
And then how you style it up.
17:04
This is becoming less of a bottleneck as it used to be,
17:08
but depending on how you dress up your context window will alter
17:14
how it behaves within the LLM.
17:16
If I use JSON, it will behave one way.
17:19
Typically, JSON has been used a lot for tool calling, so that can have
17:21
certain impacts if you do that.
17:23
Ultimately it also uses a lot more tokens to use JSON,
17:27
but you can dress it up in a lot of different ways.
17:29
We tend to use more Markdown style output using tags in order to
17:34
be able to reduce tokens, but also the LLM seems to not overemphasize
17:38
on tools when we don't use JSON.
17:45
This is an example of how we format the envelope of
17:49
information for the LLM.
17:53
All this comes back. I found this really interesting.
17:55
It came out a little bit a while ago now, but many people
17:58
have probably seen it.
17:59
At Stanford, there's a paper called ACE, Agentic Context Engineering.
18:04
And what I found really interesting about this paper is that it
18:06
is an automated version of our learning system.
18:10
So they showed that when you use agentic context engineering,
18:13
you can have this generator, reflector and curator loop that
18:16
will allow you to systematically determine ways of augmenting
18:21
your context or your prompts and your strategies inside
18:25
those prompts for potentially even gathering context in a way that
18:30
You don't need to fine-tune an LLM to see a pretty significant
18:33
uptake in terms of the actual accuracy or the metrics that
18:37
you care about in terms of the performance of your task.
18:40
And so this is very similar, when I was looking at this,
18:42
to our learning system.
18:44
The difference is that the Reflector and the
18:47
Curator is the human.
18:49
We provide the comments, the human comes in and says something
18:53
in response to those comments.
18:54
We look at that, we translate that into something we can
18:58
apply generally as a strategy going forward, and we curate
19:03
that list over time.
19:05
So it's actually, I found this really interesting because
19:06
we saw that massive improvement as a result of having that
19:09
to be personalized, but also just to do a better job overall.
19:12
And it was interesting to read a paper that kind of reinforced that.
19:18
In academia, so after each review, you know, track which
19:23
signal lead to helpful versus noisy, and this is why you
19:25
would do it if you're automating it, right, you would take
19:27
a reflecting agent, and you would say most get some patterns out
19:31
of there, say most valid issues are tied to ASG grep matches, or lint
19:35
only comments were rarely accepted.
19:37
It depends on what things you care about, right?
19:39
What are the evaluation criteria in the way that I'm going
19:42
to actually reflect on whether it's a good or bad?
19:45
And then depending on that, I would then go through and
19:48
take out some sort of learning and insert that into my original
19:52
prompt or my original prompt that leads to my context gathering.
19:57
So yeah, very interesting.
19:58
This is pretty much exactly how learnings work, only with people.
20:03
Another thing I just wanted to point out is that the way
20:06
that we handle code review, because obviously code, if you're doing
20:09
anything related to code, security is extremely important, right?
20:12
This is the IP of every company, it's the most valuable asset
20:15
outside of its people.
20:17
And so you have
20:19
You know, encrypt upon transit, ephemeral LLM queries, so
20:22
everything gets destroyed.
20:23
When we bring up a sandbox to clone the repository, as
20:26
soon as that review is done, that entire thing is destroyed.
20:28
We never store or keep any code within our actual system.
20:33
We have zero data retention and complete data isolation between
20:37
any given task, so no one else is co-located with any code base.
20:41
Because there's potential remote code execution here.
20:43
That's why our sandbox and jail environment is so strict.
20:46
Because ultimately, as there are holes in lincers or whatever
20:49
else might come up, you want to make sure that the blast
20:52
radius is contained within that single repository itself.
20:55
So you can only hurt yourself.
20:56
That's the general thesis.
21:00
I wanted to show a quick demo here.
21:06
All right.
21:08
Yeah, well, not really live, no.
21:10
Some examples.
21:13
Just generally speaking, I was going to show a few things.
21:16
So one, just to give you an idea of what the code review
21:19
could look like, and you have the ability to tune this yourself.
21:21
You can customize the summarization.
21:23
You can get rid of it. You can sort of make Code
21:26
Rabbit your own, right?
21:27
So whatever you like.
21:29
But the idea is that first thing it does is it analyzes
21:31
the PR and figures out what happened so that we can give some
21:34
level of summary so anyone coming into this has some basis of where
21:37
they are and orienting themselves to figure out what's going on.
21:41
So that's the first sort of thing.
21:42
And then we have a walkthrough which guides through each
21:45
and every file what changed and the description of that change.
21:48
So, again, anyone who comes in and wants to figure out
21:50
what's going on, I don't know if anyone's coming into a large PR and
21:53
being like, what is going on here?
21:54
This This is where you can orient yourself and figure
21:57
out what's happening.
21:59
So that comes in here and then we have a system sequence
22:02
diagram which again can help you understand the data flow within
22:05
the PR itself and if you like, there's a poem as part of this.
22:11
A few more examples of actual issues.
22:17
So here for example we have a critical issue that's caught,
22:19
I don't know if I can make Make this a little bigger.
22:22
There we go.
22:24
Code will not compile, obviously a big issue.
22:26
I'm not sure how this PR got created, right?
22:30
Invalid, await error, hand-to-suit tax.
22:32
So it's sitting there searching for all kinds of stuff, including
22:35
things that are somewhat on the obvious side of things.
22:38
But interestingly enough, you can also find complicated things
22:43
like potential race conditions.
22:46
So again, things that are inherently a little
22:48
bit even difficult for sometimes humans to find.
22:51
Concurrency issues being sort of chief among the
22:54
most difficult to understand.
22:57
And the other thing I wanted to point out for people who
22:59
use it is sometimes there are problems that we find
23:03
that are just around the PR.
23:04
So like they're in the code of the same file.
23:07
They might overlap with the PR, but ultimately they're just outside.
23:10
GitHub doesn't let us post comments there, at least not via the API.
23:14
And so we post those in this outside diff, again showing another
23:18
issue that is straddling your PR.
23:22
And ultimately, we provide a lot of patches, right, trying
23:23
to make it very easy, and on top of that, I'll show
23:26
you a little bit later, but we're actually providing a lot of prompts
23:30
for your AI systems to be able to fix these issues automatically.
23:35
And so we have another example of some critical security issues.
23:38
So someone had CodeRabbit scanned for specific types of security
23:42
issues, and we come up and we do, again, a full clone of the sandbox.
23:45
Even when you're chatting, you can actually chat with
23:47
CodeRabbit, ask questions, ask it to do extra things, and it
23:51
will actually do the entire thing.
23:52
Well, it will clone a repository, do all that stuff, agentic
23:56
loops, context gathering, to try and answer that question.
23:59
And this is a question about security vulnerabilities.
24:01
And so it went and did a thorough scan and found some issues.
24:07
Here's another one. This case, again, this was a SQL injection
24:11
vulnerability that it found.
24:13
So someone had inadvertently added something where the user
24:16
would have control over a variable that would go into a SQL query.
24:19
And so it pointed out that you might want to do some
24:22
data cleansing before that.
24:25
So all these things, finding these security issues, save
24:29
a lot in the long run of potential issues and potential data leakage.
24:33
Here's a memory leak.
24:35
That it discovered, again, this is for C and C++, and
24:40
we've all been there.
24:43
And this is the example I showed earlier, actually,
24:45
the learning where someone had done a bunch of imports and we suggested
24:48
going to star because originally they had wanted a collapse,
24:52
but that had been reverted.
24:53
And so this is an example of where the new learning
24:55
was added and the old learning was removed as a result of the
24:59
change in the way the repository wanted things to be reviewed.
25:04
Here's an example of MCP being used, so again, pulling in
25:09
context information because context is king here.
25:12
If you want to be able to find problems, you need to
25:14
know what you don't know, and so some of that involves discovery.
25:17
This is an example where MCP systems were used to discover
25:21
information about a PR.
25:25
We have this feature called pre-merge checks.
25:27
In some cases, it's not a bug in a specific part of the file.
25:31
It could be something else entirely.
25:32
And this happens after you're about, when you're ready to
25:36
merge, it'll do some other checks.
25:39
So this one is, for example, a requirement that this user
25:42
had or this repository has around feature gating.
25:45
It noticed that this new thing was not feature gated, so
25:49
the pre-merge check failed, and it gives a reason why.
25:52
And ultimately then it prevents the PR from going through
25:55
if you have it configured that way.
25:58
You can also just have it notify you and kind of let
26:00
things fall where they may, or you can have it actually reject the PR
26:04
in the same way that a human would.
26:07
And then finally, this is something relatively new.
26:12
We rolled out a planning system recently.
26:14
And so this uses the same general framework of our context gathering.
26:18
And again, all the information that we have and the way that
26:21
we view context engineering as being a longer running task to
26:24
make sure something is done well.
26:26
And this is an example of somebody had an issue that
26:28
they had brought up.
26:31
And they described the bug, the steps to reproduce the bug, and
26:34
then they asked us to plan for them and find a solution to the problem.
26:38
And so we went through, we came up with a plan, along
26:41
with props that you can go and copy into your coding
26:45
systems, so Cloud Code, or Cursor, or Codex, to implement this.
26:52
And then I wanted to just briefly show, you know, it's very easy, but
26:56
ultimately I wanted to show what it's like, what the app looks like.
27:01
And then I'll move forward.
27:04
This is more of the live demo for you.
27:07
Dangerous. All right.
27:09
So this is what a plan kind of looks like within our interfaces.
27:12
Another example of a planning system that
27:15
comes with design choices.
27:17
So these are the assumptions that were made.
27:18
Oh, it's really hard to see. It's tiny.
27:22
These are the design choices that were made by the system.
27:25
And again, the plan comes out with tasks, and you can
27:28
go through and validate them.
27:29
You can also ask for changes within the system and iterate on it.
27:33
But going back to CodeRabbit.
27:39
So you come in and you can hook up all your repositories.
27:41
We have these cool dashboards where you can see the health
27:44
of your repo as well as your organization, I should say.
27:48
All the active repositories, pull requests,
27:50
everything that's going on.
27:52
All this stuff lets you know, first of all, how many bugs am I finding
27:55
and how many bugs are my people accepting and all this other stuff.
27:58
But to let you know the health of your repository.
28:01
But yeah, that's sort of where you can configure things.
28:04
So if I go into configurations and whatnot, I can kind of
28:08
see how, first of all, I have...
28:13
An example PR that I can see what it's gonna look like
28:15
when I'm done configuring it.
28:18
But all this stuff is sort of meant to be highly adaptable.
28:21
And right now, my organization has hard settings on my PR, so
28:24
I'm not allowed to change anything.
28:25
But that's sort of the way that you get into the app.
28:29
It's really easy just clicking on, logging in with your GitHub.
28:32
All right, let's go back to here.
28:38
There's one other thing I want to play.
28:40
How do I go back?
28:52
Is it going to play? That's the question.
29:01
100% of NVIDIA's engineers use AI for code generation.
29:04
The same number of engineers check in three times the amount
29:08
of code than before.
29:09
At this scale, keeping code quality high is impossible
29:13
with just human reviewers.
29:15
CodeRabbit's review agent gives thorough feedback as volume grows.
29:19
The agent is a system of models using frontier models like
29:23
Claude and GPT for state-of-the-art capabilities and open, efficient
29:29
NVIDIA Nemotron models customized with proprietary data for
29:32
fastest response time.
29:35
CodeRabbit pulls from NVIDIA's code, knowledge base, trackers,
29:40
documentation, and more to fully understand the changes.
29:45
Using its large context, length, and reasoning capabilities,
29:48
Nemotron iteratively summarizes, extracts insights, and packages the
29:54
context for frontier models, which flag issues and suggest fixes.
29:59
As our engineers iterate, the loop continues until the code is merged.
30:05
Now, receiving feedback on code takes minutes instead of days.
30:16
Let's go back.
30:23
Let's go, all right, so these are some of the statistics we've
30:29
got from some of our customers, right, so Groupon, for example.
30:33
Not only are they finding more issues and obviously
30:35
having fewer things going into production, but actually
30:38
as you use the system, as you adopt it, and as you gain trust over what
30:41
it can do, it can actually really accelerate your ability to have
30:45
PRs merge into production, right?
30:48
So some companies, for example, used to have two mandatory
30:53
reviews on a system will end up going down to one as they learn
30:55
the system, as they personalize it and they make sure that
30:58
it does what they want it to do.
31:00
And so this obviously not only speeds things up, it
31:02
also helps alleviate some of those resources so that people can stop
31:06
trying to review the massive amount of PRs that are being created.
31:10
So 50% faster, time to first code review, number of PRs per month 36%
31:16
more, time to merge 50% faster, and number of forced merges 60% less.
31:23
So that's a good example, but ultimately the idea is here
31:26
improve velocity, ship more code.
31:27
I think that's what all of the engineers want.
31:29
We all want to build things in one level or another, regardless
31:32
of whether you are someone who's in marketing, coding, and
31:35
wanting to build something for your team or yourself to make your work
31:39
easier and make it more efficient.
31:42
Everybody wants to build something, that's why these tools are
31:44
so useful and why people are gravitating so much to them, right?
31:49
Now you have an idea, you can go out there and build it.
31:52
We want to make sure that the quality bar remains high while
31:55
that throughput is increasing.
31:58
So, a few things to remember, LLM's do reason better with
32:01
better context, but more context does not mean better context.
32:05
Domain expertise is better context, so thinking about, in this domain,
32:09
what information would I need?
32:12
So the only way to really experiment with that is to go out
32:14
and try it on some things, right?
32:16
Try and build some things, try and understand what context
32:18
is needed, and try and figure out how you can optimize that context.
32:21
I used to have an example here where I wrote a 50-line
32:26
review product in Python, and it would review some Python
32:30
code, and it came out with nothing.
32:32
And the idea was it was just looking at the diff.
32:35
And so then I took the surrounding code and put all of it in,
32:37
and it was a small enough repo you could do that.
32:40
And it would find the issue.
32:41
I think it was at that time 200,000 tokens it would use.
32:44
Seems a little bit excessive.
32:46
And then when you optimize it, the diff itself was 17,000
32:50
tokens, and by only adding 1,000 more tokens, which is
32:53
just a very carefully curated set of where it actually connected
32:57
up with the rest of the codebase, you can find the same issue.
33:00
So the idea is the optimization problem is how you end up
33:04
doing it in a cost-effective way.
33:08
So what you can do if you want to try out CodeRabbit, it's a free
33:11
trial, 14-day, you can go in and just sign up with GitHub, right?
33:13
Very easy to set it up.
33:16
And you can hook it up to JIRA Linear to understand
33:18
why you're building what you're building, and you can hook
33:21
it up very easily to your MCP systems to figure out PRD documents
33:25
and various other things like that.
33:26
So there you go.
33:29
Here's your next steps.
33:30
I want to leave some time if anybody has any questions.
33:36
Got a question right here up front.
33:41
Hi, I'm Smita and I work for NVIDIA as a program manager.
33:45
My question is, you said there's 100% data isolation and zero
33:51
data retention, also said.
33:56
You bind the context injection to optimize the token usage.
34:01
So how do you know, or how does CodeRabbit know what to bind?
34:06
How does that work? Can you double click?
34:09
Yeah so finding what you mean what to discover within the
34:12
code base right in order to understand that so every single
34:15
time we have to rediscover the context right so when
34:19
your PR comes in it's connected to certain information so we
34:22
know certain things are going to be needed so we look at the code graph
34:25
so we actually build at that moment
34:28
A connection point between your code and the ways that it sort of
34:31
connects out from a graph, right?
34:33
So if I'm using a function from a different class, that
34:35
would be an edge, right?
34:37
If I'm importing certain things, that would be another edge.
34:40
And so I build a graph.
34:41
And I can build that a certain number of layers deep, and then
34:44
I know that's a starting point.
34:46
And based on that, I now have additional contexts.
34:48
I don't have to put all the code in there, but I have
34:51
some more information.
34:53
And now I can use that to do agentic discovery, where
34:55
I can actually start to...
34:57
Run shell scripts and bash scripts, ripgrep, things like
35:00
this to discover additional places I might need to look at.
35:04
And this is what happens sort of in the background before the
35:06
review event actually takes place.
35:09
Every single review, that happens again.
35:12
And every PR is slightly different, right?
35:13
So we don't need the same information each time.
35:16
So every time that could be large, depending if you have a large PR,
35:20
or it could actually be very small if you're changing very little,
35:22
or your changes are self-contained within the file, for example.
35:27
Very cool. I do want to remind people, we do have a raffle.
35:30
I'm going to come to the back.
35:31
I appreciate everyone who sits up front, but nobody in the
35:33
back ever gets to ask questions.
35:39
Hey, Max Gerardsen.
35:41
Great presentation, by the way.
35:43
Really interesting.
35:44
I had a question.
35:46
Earlier in your talk, you discussed how you can actually
35:51
have a conversation with the CodeRabbit and sort of direct
35:56
it into what your company best practices are, right?
36:01
And so you can do that at the local level and also at the CI level.
36:06
And is there communication from one level to the other and how would
36:12
you manage sort of conflicting claims, especially at the CI level?
36:17
So usually we try and detect conflicting claims
36:19
when they come in.
36:20
So if someone's chatting with CodeRabbit in the GitHub
36:25
ecosystem, so you saw before we deleted an old learning.
36:30
Based on the new learning because we detected that they
36:32
overlapped, right?
36:33
So we're going to show conflicts that happen as they happen
36:37
and we will update you in terms of how learnings either have
36:40
been added, removed, or updated.
36:42
So if you do something that's slightly related, we might
36:44
actually build on top of an existing learning similar
36:47
to how the ACE paper works.
36:50
But yeah, conflicts are something that you have to manage, right?
36:52
So something you have to actively look for.
# Practical Context Engineering: Eliminate Bugs With High-Signal AI Code Reviews
David Loker,VP of AI,CodeRabbit
Harjot Gill,CEO,CodeRabbit
AI-Generated Summary of this Video
Rate Now
AI-assisted coding has sped up code generation, but human-only code reviews are now the bottleneck. Too many pull requests (PRs), not enough reviewer bandwidth. This session shows how you can build the right context-engineering architecture that helps LLMs deliver high-signal AI code reviews.
Catch logical or functional bugs, memory leaks, security vulnerabilities, code refactors, edge cases, and more. You'll see how bringing in the right context from external datasets (such as code graphs, PR history, issue tickets, MCP servers, and linters) yields better-quality reviews of PR diffs. You will leave with a practical playbook to pilot and scale AI code reviews that increase your release velocity with fewer bugs.
### Learn More About This Topic
Share
Favorite
Add to list
PDF
Events & Trainings:GTC San Jose
Date:March 2026
Topic:Agentic AI / Generative AI - Code / Software Generation
Industry:All Industries
Level:General Interest
Language:English
NVIDIA technology:NVIDIA NIM
Region:
Company Information
* [About Us](https://www.nvidia.com/en-us/about-nvidia/)
* [Investors](https://investor.nvidia.com/home/default.aspx)
* [Venture Capital (NVentures)](https://www.nvidia.com/en-us/startups/nventures/)
* [NVIDIA Foundation](https://www.nvidia.com/en-us/foundation/)
* [Research](https://www.nvidia.com/en-us/research/)
* [Corporate Sustainability](https://www.nvidia.com/en-us/sustainability/)
* [Technologies](https://www.nvidia.com/en-us/technologies/)
* [Careers](https://www.nvidia.com/en-us/about-nvidia/careers/)
News and Events
* [Newsroom](https://nvidianews.nvidia.com/)
* [Company Blog](https://blogs.nvidia.com/)
* [Technical Blog](https://developer.nvidia.com/blog/)
* [Webinars](https://www.nvidia.com/en-us/about-nvidia/webinar-portal/)
* [Stay Informed](https://www.nvidia.com/en-us/preferences/email-signup/)
* [Events Calendar](https://www.nvidia.com/en-us/events/)
* [GTC AI Conference](https://www.nvidia.com/gtc/events/)
* [NVIDIA On-Demand](https://www.nvidia.com/en-us/on-demand/)
Popular Links
* [Developers](https://developer.nvidia.com/)
* [Partners](https://www.nvidia.com/en-us/about-nvidia/partners/)
* [Executive Insights](https://www.nvidia.com/en-us/executive-insights/)
* [Startups and VCs](https://www.nvidia.com/en-us/startups/)
* [Documentation](https://docs.nvidia.com/)
* [Technical Training](https://www.nvidia.com/en-us/learn/organizations/)
* [Professional Services for Data Science](https://www.nvidia.com/en-us/support/enterprise/advisory-services/)
Follow NVIDIA
[](https://www.facebook.com/NVIDIA "")[](https://www.instagram.com/nvidia/?hl=en)[](https://www.linkedin.com/company/nvidia/)[](https://twitter.com/nvidia "")[](https://www.youtube.com/user/nvidia)
[United States](https://www.nvidia.com/en-us/location-selector/)
* [Privacy Policy](https://www.nvidia.com/en-us/about-nvidia/privacy-policy/)
* [Your Privacy Choices](https://www.nvidia.com/en-us/about-nvidia/privacy-center/)
* [Terms of Service](https://www.nvidia.com/en-us/about-nvidia/terms-of-service/)
* [Accessibility](https://www.nvidia.com/en-us/about-nvidia/accessibility/)
* [Corporate Policies](https://www.nvidia.com/en-us/about-nvidia/company-policies/)
* [Product Security](https://www.nvidia.com/en-us/product-security/)
* [Contact](https://www.nvidia.com/en-us/contact/)
Copyright © 2026 NVIDIA Corporation
Select Location
The Americas
* [Argentina](https://www.nvidia.com/es-la/ "Argentina")
* [Brasil (Brazil)](https://www.nvidia.com/pt-br/ "Brasil (Brazil)")
* [Canada](https://www.nvidia.com/en-us/ "Canada")
* [Chile](https://www.nvidia.com/es-la/ "Chile")
* [Colombia](https://www.nvidia.com/es-la/ "Colombia")
* [México (Mexico)](https://www.nvidia.com/es-la/ "México (Mexico)")
* [Peru](https://www.nvidia.com/es-la/ "Peru")
* [United States](https://www.nvidia.com/en-us/ "United States")
Europe
* [België (Belgium)](https://www.nvidia.com/nl-nl/ "België (Belgium)")
* [Belgique (Belgium)](https://www.nvidia.com/fr-be/ "Belgique (Belgium)")
* [Česká Republika (Czech Republic)](https://www.nvidia.com/cs-cz/ "Česká Republika (Czech Republic)")
* [Danmark (Denmark)](https://www.nvidia.com/da-dk/ "Danmark (Denmark)")
* [Deutschland (Germany)](https://www.nvidia.com/de-de/ "Deutschland (Germany)")
* [España (Spain)](https://www.nvidia.com/es-es/ "España (Spain)")
* [France](https://www.nvidia.com/fr-fr/ "France")
* [Italia (Italy)](https://www.nvidia.com/it-it/ "Italia (Italy)")
* [Nederland (Netherlands)](https://www.nvidia.com/nl-nl/ "Nederland (Netherlands)")
* [Norge (Norway)](https://www.nvidia.com/nb-no/ "Norge (Norway)")
* [Österreich (Austria)](https://www.nvidia.com/de-at/ "Österreich (Austria)")
* [Polska (Poland)](https://www.nvidia.com/pl-pl/ "Polska (Poland)")
* [România (Romania)](https://www.nvidia.com/ro-ro/ "România (Romania)")
* [Suomi (Finland)](https://www.nvidia.com/fi-fi/ "Suomi (Finland)")
* [Sverige (Sweden)](https://www.nvidia.com/sv-se/ "Sverige (Sweden)")
* [Türkiye (Turkey)](https://www.nvidia.com/tr-tr/ "Türkiye (Turkey)")
* [United Kingdom](https://www.nvidia.com/en-gb/ "United Kingdom")
* [Rest of Europe](https://www.nvidia.com/en-eu/ "Rest of Europe")
Asia
* [Australia](https://www.nvidia.com/en-au/ "Australia")
* [中国大陆 (Mainland China)](https://www.nvidia.com/zh-cn/ "中国大陆 (Mainland China)")
* [India](https://www.nvidia.com/en-in/ "India")
* [日本 (Japan)](https://www.nvidia.com/ja-jp/ "日本 (Japan)")
* [대한민국 (South Korea)](https://www.nvidia.com/ko-kr/ "대한민국 (South Korea)")
* [Singapore](https://www.nvidia.com/en-sg/ "Singapore")
* [台灣 (Taiwan)](https://www.nvidia.com/zh-tw/ "台灣 (Taiwan)")
Middle East
* [Middle East](https://www.nvidia.com/en-me/ "Middle East")
NVIDIA uses cookies to improve your experience on our web site. We and our third-party partners also use cookies and other tools to collect and record information you provide as well as information about your interactions with our websites for performance improvement, analytics, and to assist in marketing efforts. By clicking "Accept All", you consent to our use of cookies and other tools as described in our [Cookie Policy](https://www.nvidia.com/en-us/about-nvidia/cookie-policy/). You can manage your cookie settings by clicking on "Manage Settings." By continuing to use this site or by clicking one of the buttons below, you agree to our [Terms of Service](https://www.nvidia.com/en-us/about-nvidia/terms-of-service/) (which contains important waivers). Please see our [Privacy Policy](https://www.nvidia.com/en-us/about-nvidia/privacy-policy/) for more information on our privacy practices.
We have detected the Global Privacy Control (GPC) signal and recorded your rejection of all optional cookies on this site for this browser. You can manage your cookie settings by clicking on "Manage Settings". Please see our [Cookie Policy](https://www.nvidia.com/en-us/about-nvidia/cookie-policy/) for more information. To opt out of non-cookie personal information "sales" / "sharing" for targeted advertising purposes, please visit the [NVIDIA Preference Center](https://www.nvidia.com/en-us/about-nvidia/privacy-center/). Please see our [Privacy Policy](https://www.nvidia.com/en-us/about-nvidia/privacy-policy/) for more information on our privacy practices.
We have detected the Global Privacy Control Signal (GPC) and have opted you out of all optional cookies on this browser. You can manage your cookie settings by clicking on "Manage Settings". Please see our [Cookie Policy](https://www.nvidia.com/en-us/about-nvidia/cookie-policy/) for more information. We have also opted you out of "sharing"/"sales" of personal information outside of cookies. You can manage these settings in the NVIDIA [NVIDIA Preference Center](https://www.nvidia.com/en-us/privacy-center/). Please see our [Privacy Policy](https://www.nvidia.com/en-us/about-nvidia/privacy-policy/) for more information.
We have detected the Global Privacy Control Signal (GPC) and have opted you out of all optional cookies on this browser. You can manage your cookie settings by clicking on "Manage Settings". Please see our [Cookie Policy](https://www.nvidia.com/en-us/about-nvidia/cookie-policy/) for more information. We have also opted you out of "sharing"/"sales" of personal information outside of cookies which overrides at least one of your previous settings. You can manage them in the [NVIDIA Preference Center](https://www.nvidia.com/en-us/privacy-center/). Please see our [Privacy Policy](https://www.nvidia.com/en-us/about-nvidia/privacy-policy/) for more information.
Manage Settings
Reject Optional Accept All
Cookie Settings
We and our third-party partners (including social media, advertising, and analytics partners) use cookies and other tracking technologies to collect, store, monitor, and process certain information about you when you visit our website. The information collected might relate to you, your preferences, or your device. We use that information to make the site work, analyze performance and traffic on our website, provide a more personalized web experience, and assist in our marketing efforts.
Under certain privacy laws, you have the right to direct us not to "sell" or "share" your personal information for targeted advertising. To opt-out of the "sale" and "sharing" of personal information through cookies, you must opt-out of optional cookies using the toggles below. To opt out of the "sale" and "sharing" of data collected by other means (e.g., online forms) you must also update your data sharing preferences through the [NVIDIA Preference Center](https://www.nvidia.com/en-us/about-nvidia/privacy-center/).
Click on the different category headings below to find out more and change the settings according to your preference. You cannot opt out of Required Cookies as they are deployed to ensure the proper functioning of our website (such as prompting the cookie banner and remembering your settings, etc.). By clicking "Save and Accept" or "Decline All" at the bottom, you consent to the use of cookies and other tools as described in our [Cookie Policy](https://www.nvidia.com/en-us/about-nvidia/cookie-policy/) in accordance with your settings and accept our [Terms of Service](https://www.nvidia.com/en-us/about-nvidia/terms-of-service/) (which contains important waivers). For more information about our privacy practices, please see our [Privacy Policy](https://www.nvidia.com/en-us/about-nvidia/privacy-policy/).
Required Cookies
Always Active
These cookies enable core functionality such as security, network management, and accessibility. These cookies are required for the site to function and cannot be turned off.
Cookies Details
Performance Cookies
- [x] Performance Cookies
These cookies are used to provide quantitative measures of our website visitors, such as the number of times you visit, time on page, your mouse movements, scrolling, clicks and keystroke activity on the websites; other browsing, search, or product research behavior; and what brought you to our site. These cookies may store a unique ID so that our system will remember you when you return. Information collected with these cookies is used to measure and find ways to improve website performance.
Cookies Details
Personalization Cookies
- [x] Personalization Cookies
These cookies collect data about how you have interacted with our website to help us improve your web experience, such as which pages you have visited. These cookies may store a unique ID so that our system will remember you when you return. They may be set by us or by third party providers whose services we have added to our pages. These cookies enable us to provide enhanced website functionality and personalization as well as make the marketing messages we send to you more relevant to your interests. If you do not allow these cookies, then some or all of these services may not function properly.
Cookies Details
Advertising Cookies
- [x] Advertising Cookies
These cookies record your visit to our websites, the pages you have visited and the links you have followed to influence the advertisements that you see on other websites. These cookies and the information they collect may be managed by other companies, including our advertising partners, and may be used to build a profile of your interests and show you relevant advertising on other sites. We and our advertising partners will use this information to make our websites and the advertising displayed on it, more relevant to your interests.
Cookies Details
Cookie List
Clear
* - [x] checkbox label label
Apply Cancel
Consent Leg.Interest
- [x] checkbox label label
- [x] checkbox label label
- [x] checkbox label label
Decline All Save and Accept
[](https://www.onetrust.com/solutions/consent-and-preferences/)
Copy debug info