blocks|key|2222691|text|type|unstyled|depth|inlineStyleRanges|entityRanges|data|entityMap^0^^$0|@$1|2|3|-4|4|5|6|B|7|@]|8|@]|9|$]]]|A|$]]

<p>You may not need that setting and it may be worth knowing what &quot;low latency&quot; means in detail to see how it adjust to whatever you need to test:</p>
<h1>Kernel &quot;low latency&quot;</h1>
<p>When the kernel is built in low latency, it ticks at 1000hz</p>
<p>This means the kernel will preempt processes 1000 times per second as opposed to 100-200 times (or Windows' maximum 1000hz when <a href="https://docs.microsoft.com/en-us/windows/win32/api/timeapi/nf-timeapi-timebeginperiod" rel="nofollow noreferrer">timeBeginPeriod(1)</a> is called)</p>
<p>The details are explained in <a href="https://unix.stackexchange.com/a/553993/350793">another post</a> but basically it boils down to context switching happening more often which means sleeping threads can wake faster and things that need to be processed ASAP (like an audio server) are woken up earlier.</p>
<p>It basically makes the kernel check on each process &quot;are you done yet?&quot; much more often.</p>
<p>The cost is higher CPU overhead from waking up so often, and horrible energy efficiency (btw you should keep your CPU governor in performance for stable readings)</p>
<h1>Graphics &quot;low latency&quot;</h1>
<p>Low latency in graphics is defined as &quot;the time it takes for an input, like a keyboard stroke, to appear on screen&quot;.</p>
<h2>VSync off</h2>
<p>When VSync is off, you already get the lowest latency; because if graphics are rendered at 300 fps on a 60hz monitor; it will be shown at 300hz and be displayed immediately even if the monitor is still displaying the previous frame, which causes <a href="https://www.youtube.com/watch?v=8ENCV4xUjj0" rel="nofollow noreferrer">visible tearing</a> as two or more frames overlap.</p>
<p>There is a limit on how fast it can be shown on screen even if VSync is off, which will be limited by the monitor's pixel response (usually 5ms) and other HW limits.</p>
<p>NVIDIA settings don't matter here, <strong>because you're already getting the lowest possible latency</strong>.</p>
<h2>VSync On</h2>
<p>VSync waits for the VBlank interval in order to swap the image. This means you get a perfect picture on the monitor.</p>
<p>However most monitors display at 60hz (unless you buy e.g. a 240hz one).</p>
<p>If GPU was still working on a frame when the VBlank interval came, it needs to wait another 16.666ms (I'll truncate to 16ms for brevity) so it can be displayed. This is known as &quot;missing&quot; the VBlank interval, and thus latency is added.</p>
<p>Another problem is that if your app takes e.g. 2ms to process and started right after VBlank, now you need to sit idle for 14ms before it can be displayed on screen.</p>
<p>By the time it gets displayed on screen, the input being shown was gathered at least 16ms ago or more (assuming your app started right after VBlank, if it started 8ms after VBlank, then it took you 2ms to process, 6ms of idle, and the inputs are at least 8ms old or more).</p>
<p>This gets exacerbated with triple buffering which may add <em>up to</em> a whole frame of latency (16ms!)</p>
<p>NVIDIA's low latency option is a combination of techniques and hacks (often tuned for the popular games, which means it may not always work in your particular application) that when enabled:</p>
<ol>
<li>Double Buffering is favoured over Triple Buffer. I believe this option is already available on NVIDIA settings separately.</li>
<li>It <strong>forces the app to sleep</strong>. We said in your example that you take 2ms to process (I'm making the numbers up, it's an example). If this is stable, the ideal is to sleep for roughly 13ms after vblank, gather your buffered input (e.g. keyboard, sensors, etc), process everything in 2ms, and you still have 1ms of spare for overhead and display on screen. Hence the image displayed on screen shows the inputs roughly gathered 3ms ago with no tearing (+5ms of pixel response so it's probably 8ms ago).
<ul>
<li>The tricky part is knowing when vblank happens</li>
<li>The even more tricky part is estimating how long you will take to process inputs. It's rarely a perfect 2ms and there may be noise and jitter (OS allocations will destroy you). If you overestimate you'll miss the next Vblank and will have to wait another 16ms</li>
<li>For more info see <a href="https://www.gdcvault.com/play/1026327/Controller-to-Display-Latency-in" rel="nofollow noreferrer">Controller to Display Latency in Call of Duty</a></li>
</ul>
</li>
</ol>
<h1>What this means</h1>
<ol>
<li>I didn't mention this, but Python is a terrible choice for low latency stuff. It's an interpreted language with lots of memory allocations which causes a lot of jitter. That's why audio servers are written in C (and not just any C, you have to adhere to some rules like avoid most mutex or use them sparingly, do not call malloc, etc). That depends on how extreme the low latency is. Having a strict 3ms requirement isn't the same as a loose 20ms requirement.</li>
<li>You can turn off Triple Buffering in NVIDIA settings.</li>
<li>If you don't care about tearing, turn VSync off. You're done.</li>
<li>You can solve all the next problems with money, by buying a 240hz monitor. You're done.</li>
<li>If you care about tearing and can't buy a 240hz monitor, prefer using Vulkan (<em>or an engine that uses Vulkan</em> and exposes to you settings such as triple/double buffer and mailbox presentation) which needs more expert knowledge but gives you more tools to know when VBlank happens and how presentation happens (e.g. FIFO vs Mailbox presentation; I won't go into detail but you'll want to use Mailbox). You are also in full control of whether Double or Triple buffer is used.</li>
<li>Take measurements and test how long you need to <code>sleep()</code> on a given machine. Once you found the sweet spot this may be 'good enough'.</li>
<li>Ideally, use DRM instead of Xorg or Wayland interface (this is VERY advanced) to control the screen directly.</li>
<li>If you only care about generating a report and analyzing/crunching numbers later; and displaying on screen in real time as fast as possible is not necessary; then this entire section has been a waste of time because &quot;NVIDIA low latency&quot; settings won't affect your readings; it only controls how old the data was when it's displayed on screen.</li>
</ol>
<h1>Other misc</h1>
<p>Set CPU governor to performance for stable readings. Disable the GPU's power saving for the same reason (how to do that varies on vendor / driver).</p>
<p><a href="https://www.youtube.com/watch?v=EtLHLfNpu84" rel="nofollow noreferrer">Fighting Latency on Call of Duty: Black Ops III</a> is a recommended reading on how they setup a GoPro camera to identify sources of latency and fix it; even though most of the talk is oriented towards multiplayer network latency (which is in the order of 30-200ms)</p>
<p>Ultimately low latency is a deep topic and I'm afraid a single toggle may or may not do what you want; depending on what you need.</p>


<p>I should run a visual experiment project with Matlab or Python in which the time resolution is dramatically important! that's why I shift to Linux and use the low-latency kernel in ubuntu. but I thought there is an ultra-low-latency or something like that in Linux Nvidia settings, same as a thing in 3D acceleration in Nvidia control panel windows.</p>
<p>Is there any way that enables Nvidia ultra-low latency in ubuntu 22.04?</p>


Ultra low-latency in ubuntu 22.04

翻译质量差，导致语言生硬或混乱。

没有提供实际的解决方法或示例。

解答不清晰，无法理解或解决问题。

页面排版不美观，阅读体验差。

文章

问答

视频

教程

学习中心

腾讯云实验室

直播

竞赛

腾讯云代码分析专区

腾讯iOA零信任安全管理系统专区

腾讯云架构师技术同盟交流圈

腾讯云数据库专区

腾讯云智能顾问专区

腾讯云原生专区

腾讯混元专区

腾讯云TCE专区

腾讯云Lighthouse专区

腾讯云HAI专区

腾讯云Edgeone专区

腾讯云存储专区

腾讯云智能专区

腾讯轻联专区 

腾讯云开发专区

TAPD专区

腾讯轻量云游戏服专区

腾讯云最具价值专家

腾讯云架构师技术同盟

腾讯云创作之星

腾讯云开发者先锋

腾讯云AI代码助手

云原生构建

TAPD 敏捷项目管理

Cloud Studio

SDK中心

API中心

命令行工具

功能1上新10个字符

功能2描述100个字符功能2描述100个字符功能2描述100个字符功能2描述100个字符功能2描述100个字符功能2描述100个字符功能2描述100个字符功能2描述100个字符功能2描述100个字符。

功能2上新100个字符功能2上新100个字符功能2上新100个字符功能2上新100个字符功能2上新100个字符功能2上新100个字符功能2上新100个字符功能2上新100个字符功能2上新100个字符。

功能5描述100个字符功能5描述100个字符功能5描述100个字符功能5描述100个字符功能5描述100个字符功能5描述100个字符

功能5上新100个字符功能5上新100个字符功能5上新100个字符功能5上新100个字符功能5上新100个字符功能5上新100个字符功能5上新100个字符功能5上新100个字符功能5上新100个字符功能5上新100个字符

功能4上新

文章&问答评论现已支持表情

全新交互，全新视觉，新增快捷键、悬浮工具栏、高亮块等功能并同时优化现有功能，全面提升创作效率和体验

社区富文本编辑器全新改版！诚邀体验～ 

精选全网热门MCP server，让你的AI更好用 🚀

💥开发者 MCP广场重磅上线！

涵盖代码开发、场景应用、自动测试全流程，助你从零构建专属AI助手

一站式MCP教程库，解锁AI应用新玩法

聚焦“写作效率、视觉美观与运行性能”三方面进行全面升级，为您提供更高效、稳定的创作环境

社区富文本&Markdown编辑器全新改版上线，欢迎大家体验!

诚挚邀请您参与本次调研，分享您的真实使用感受与建议。您的反馈至关重要，感谢您的支持与参与！

社区新版编辑器体验调研

我应该使用Matlab或Python运行一个可视化实验项目，其中时间分辨率非常重要！这就是为什么我转向Linux，在ubuntu中使用低延迟内核的原因。但我认为在Linux Nvidia设置中有一个超低延迟或类似的东西，就像Nvidia控制面板窗口中的3D加速一样。是否有任何方法可以使Nvidia超低延迟在ubuntu 22.04？

问ubuntu 22.04的超低延迟
EN

回答 1

Ask Ubuntu用户

社区

活动

圈层

关于

腾讯云开发者

热门产品

热门推荐

更多推荐

问ubuntu 22.04的超低延迟EN