Performance profiling tips for game developers

Smooth performance is essential to creating immersive gaming experiences for players. By profiling and honing your game’s performance for a broad range of platforms and devices, you can expand your player base and increase your chance for success.

This page outlines a general profiling workflow for game developers. It’s excerpted from the e-book, Ultimate guide to profiling Unity games, available to download for free. The e-book was created by both external and internal Unity experts in game development, profiling, and optimization.

Read on to learn about useful goals to set with profiling, common performance bottlenecks, such as being CPU-bound or GPU-bound, and how to identify and investigate these situations in more detail.

FPS 프레임 속도 그래프

Set a frame budget

Measuring your game’s frame rate in frames per second (fps) is not ideal for delivering consistent experiences for your players. Consider the following simplified scenario:

During runtime, your game renders 59 frames in 0.75 seconds. However, the next frame takes 0.25 seconds to render. The average delivered frame rate of 60 fps sounds good, but in reality players will notice a stutter effect since the last frame takes a quarter of a second to render.

This is one of the reasons why it’s important to aim for a specific time budget per frame. This  provides you with a solid goal to work toward when profiling and optimizing your game, and ultimately, it creates a smoother and more consistent experience for your players.

Each frame will have a time budget based on your target fps. An application targeting 30 fps should always take less than 33.33 ms per frame (1000 ms / 30 fps). Likewise, a target of 60 fps leaves 16.66 ms per frame  (1000 ms / 60 fps).

You can exceed this budget during non-interactive sequences, for example, when displaying UI menus or scene loading, but not during gameplay. Even a single frame that exceeds the target frame budget will cause hitches. 

Note: A consistently high frame rate in VR games is essential to avoid causing nausea or discomfort to players. Without it, you risk being rejected by the platform holder during your game’s certification.

Frames per second: A deceptive metric

A common way that gamers measure performance is with frame rate, or frames per second. However, it’s recommended that you use frame time in milliseconds instead. To understand why, look at the above graph of fps versus frame time. 

Consider these numbers:

     1000 ms/sec / 900 fps = 1.111 ms per frame
     1000 ms/sec / 450 fps = 2.222 ms per frame

     1000 ms/sec / 60 fps = 16.666 ms per frame
     1000 ms/sec / 56.25 fps = 17.777 ms per frame

If your application is running at 900 fps, this translates into a frame time of 1.111 milliseconds per frame. At 450 fps, this is 2.222 milliseconds per frame. This represents a difference of only 1.111 milliseconds per frame, even though the frame rate appears to drop by one half.

If you look at the differences between 60 fps and 56.25 fps, that translates into 16.666 milliseconds per frame and 17.777 milliseconds per frame, respectively. This also represents 1.111 milliseconds extra per frame, but here, the drop in frame rate feels far less dramatic percentage-wise.

This is why developers use the average frame time to benchmark game speed rather than fps. 

Don’t worry about fps unless you drop below your target frame rate. Focus on frame time to measure how fast your game is running, then stay within your frame budget.

Read the original article, “Robert Dunlop’s fps versus frame time,” for more information.

모바일 도전 과제

FTrace 또는 Perfetto와 같은 툴로 CPU 주파수와 대기 상태를 모니터링하면 프레임 예산 비용 최적화의 결과를 식별할 수 있습니다.

Mobile challenges

Thermal control is one of the most important areas to optimize for when developing applications for mobile devices. If the CPU or GPU spend too long working at full throttle due to inefficient design, those chips will get hot. To avoid damage to the chips (and potentially burning a player’s hands!), the operating system will reduce the clock speed of the device to allow it to cool down, causing frame stuttering and a poor user experience. This performance reduction is known as thermal throttling.

Higher frame rates and increased code execution (or DRAM access operations) lead to increased battery drain and heat generation. Bad performance can also cut out entire segments of lower-end mobile devices, which can lead to missed market opportunities, and therefore, lower sales.

When taking on the problem of thermals, consider the budget you have to work with as a system-wide budget.

Combat thermal throttling and battery drain by leveraging an early profiling technique to optimize your game from the start. Dial in your project settings for your target platform hardware to combat thermal and battery drain problems.

Adjust frame budgets on mobile

Leaving a frame idle time of around 35% is the typical recommendation to combat device thermal issues over extended play times. This gives mobile chips time to cool down and helps to prevent excessive battery drain. Using a target frame time of 33.33 ms per frame (for 30 fps), a typical frame budget for mobile devices will be approximately 22 ms per frame.

The calculation looks like this: (1000 ms / 30) * 0.65 = 21.66 ms

To achieve 60 fps on mobile using the same calculation would require a target frame time of (1000 ms / 60) * 0.65 = 10.83 ms. This is difficult to achieve on many mobile devices and would drain the battery twice as fast as targeting 30 fps. For these reasons, most mobile games target 30 fps rather than 60. Use Application.targetFrameRate to control this setting, and refer to the “Set a frame budget” section in the e-book for more details about frame time.

Frequency scaling on mobile chips can make it tricky to identify your frame idle time budget allocations when profiling. Your improvements and optimizations can have a net positive effect, but the mobile device might be scaling frequency down, and as a result, running cooler. Use custom tooling such as FTrace or Perfetto to monitor mobile chip frequencies, idle time, and scaling before and after optimizations.

As long as you stay within your total frame time budget for your target fps (33.33 ms for 30 fps) and see your device working less or logging lower temperatures to maintain this frame rate, then you’re on the right track.

Another reason to add breathing room to frame budget on mobile devices is to account for real-world temperature fluctuations. On a hot day, a mobile device will heat up and have trouble dissipating heat, which can lead to thermal throttling and poor game performance. Setting aside a percent of the frame budget will help to avoid these sorts of scenarios.

메모리 액세스 작업 줄이기

Arm의 Streamline 성능 분석기에는 대상 Arm 하드웨어에서 라이브 프로파일링 세션 동안 캡처할 수 있는 풍부한 성능 카운터 정보가 포함되어 있습니다.

Reduce memory access operations

DRAM access is typically a power-hungry operation on mobile devices. Arm’s optimization advice for graphics content on mobile devices says that LPDDR4 memory access costs approximately 100 picojoules per byte. 

Reduce the number of memory access operations per frame by:

  • Reducing frame rate
  • Reducing display resolution where possible
  • Using simpler meshes with reduced vertex count and attribute precision
  • Using texture compression and mipmapping

When you need to focus on devices leveraging Arm or Arm Mali hardware, Arm Mobile Studio tooling (specifically, Streamline Performance Analyzer) includes some great performance counters for identifying memory bandwidth issues. The counters are listed and explained for each Arm GPU generation, for example, Mali-G78. Note that Mobile Studio GPU profiling requires Arm Mali.

Establish hardware tiers for benchmarking

In addition to using platform-specific profiling tools, establish tiers or a lowest-spec device for each platform and tier of quality you wish to support, then profile and optimize performance for each of these specifications.

As an example, if you’re targeting mobile platforms, you might decide to support three tiers with quality controls that toggle features on or off based on the target hardware. You then optimize for the lowest device specification in each tier. As another example, if you’re developing a game for both PlayStation 4 and PlayStation 5, make sure you profile on both.

For a complete mobile optimization guide, take a look at Optimize your mobile game performance. This e-book has many tips and tricks that will help you reduce thermal throttling and increase battery life for mobile devices running your games.

From high- to low-level profiling

A top-to-bottom approach works well when profiling, starting with Deep Profiling disabled. Use this high-level approach to collect data and take notes on which scenarios cause unwanted managed allocations or too much CPU time in your core game loop areas.

You’ll need to first gather call stacks for GC.Alloc markers. If you’re unfamiliar with this process, find some tips and tricks in the “Locating recurring memory allocations over application lifetime” section in Ultimate guide to profiling Unity games.

If the reported call stacks are not detailed enough to track down the source of the allocations or other slowdowns, you can then perform a second profiling session with Deep Profiling enabled in order to find the source of the allocations.

When collecting notes on the frame time “offenders,” be sure to note how they compare relative to the rest of the frame. This relative impact will be affected by turning on Deep Profiling.

Read more about deep profiling in Ultimate guide to profiling Unity games.

Profile early

The best gains from profiling are made when you start early on in your project’s development lifecycle.

Profile early and often so you and your team understand and memorize a “performance signature” for the project. If performance takes a nosedive, you’ll be able to easily spot when things go wrong and remedy the issue.

The most accurate profiling results always come from running and profiling builds on target devices, together with leveraging platform-specific tooling to dig into the hardware characteristics of each platform. This combination will provide you with a holistic view of application performance across all your target devices.

프로파일링 플로우 차트

Identify performance issues

Download the printable PDF version of this chart here.

On some platforms, determining whether your application is CPU- or GPU-bound is easy. For example, when running an iOS game from Xcode, the fps panel shows a bar chart with the total CPU and GPU time so you can see which is the highest. The CPU time includes time spent waiting for VSync, which is always enabled on mobile devices.

However, on some platforms it can be challenging to get GPU timing data. Fortunately, the Unity Profiler shows enough information to identify the location of performance bottlenecks. The flow chart above illustrates the initial profiling process with the sections following it providing detailed information on each step. They also present Profiler captures from real Unity projects to illustrate the kinds of things to look for.

To get a holistic picture of all CPU activity, including when it’s waiting for the GPU, use the Timeline view in the CPU module of the Profiler. Familiarize yourself with the common Profiler markers to interpret captures correctly. Some of the Profiler markers may appear differently depending on your target platform, so spend time exploring captures of your game on each of your target platforms to get a feel for what a “normal” capture looks like for your project.

A project’s performance is bound by the chip and/or thread that takes the longest. That’s the area where you should focus your optimization efforts. For example, imagine a game with a target frame time budget of 33.33 ms and VSync enabled:

  • If the CPU frame time (excluding VSync) is 25 ms and GPU time is 20 ms, no problem! You’re CPU-bound, but everything is within budget, and optimizing things won’t improve the frame rate (unless you get both CPU and GPU below 16.66 ms and jump up to 60 fps).
  • If the CPU frame time is 40 ms and GPU is 20 ms, you’re CPU-bound and will need to optimize the CPU performance. Optimizing the GPU performance won’t help; in fact, you might want to move some of the CPU work onto the GPU, for example, by using compute shaders instead of C# code for some things, to balance things out.
  • If the CPU frame time is 20 ms and GPU is 40 ms, you’re GPU-bound and need to optimize the GPU work.
  • If CPU and GPU are both at 40 ms, you’re bound by both and will need to optimize both below 33.33 ms to reach 30 fps.

See these resources that further explore being CPU- or GPU-bound:

프레임 예산에 맞춰야 하나요?

과열 없이 30fps를 구현하기 위해 필요한 최대 22ms 프레임 예산 내에서 원활하게 작동하는 게임입니다.

Are you within frame budget?

Profiling and optimizing your project early and often throughout development will help you ensure that all of your application’s CPU threads and the overall GPU frame time are within the frame budget.  

Above is an image of a Profiler capture from a Unity mobile game developed by a team that did ongoing profiling and optimization. The game targets 60 fps on high-spec mobile phones, and 30 fps on medium/low-spec phones, such as the one in this capture.

Note how nearly half of the time on the selected frame is occupied by the yellow WaitForTargetfps Profiler marker. The application has set Application.targetFrameRate to 30 fps, and VSync is enabled. The actual processing work on the main thread finishes at around the 19 ms mark, and the rest of the time is spent waiting for the remainder of the 33.33 ms to elapse before beginning the next frame. Although this time is represented with a Profiler marker, the main CPU thread is essentially idle during this time, allowing the CPU to cool and using a minimum of battery power.

The marker to look out for might be different on other platforms or if VSync is disabled. The important thing is to check whether the main thread is running within your frame budget or exactly on your frame budget, with some kind of marker that indicates that the application is waiting for VSync and whether the other threads have any idle time. 

Idle time is represented by gray or yellow Profiler markers. The screenshot above shows that the render thread is idling in Gfx.WaitForGfxCommandsFromMainThread, which indicates times when it has finished sending draw calls to the GPU on one frame, and is waiting for more draw call requests from the CPU on the next. Similarly, although the Job Worker 0 thread spends some time in Canvas.GeometryJob, most of the time it’s Idle. These are all signs of an application that’s comfortably within the frame budget.

If your game is in frame budget

If you are within the frame budget, including any adjustments made to the budget to account for battery usage and thermal throttling, you have finished performance profiling until next time – congratulations. Consider running the Memory Profiler to ensure that the application is also within its memory budget.

The image above shows a game running comfortably within the ~22 ms frame budget required for 30 fps. Note the WaitForTargetfps padding the main thread time until VSync and the gray idle times in the render thread and worker thread. Also note that the VBlank interval can be observed by looking at the end times of Gfx.Present frame over frame, and that you can draw up a time scale in the Timeline area or on the Time ruler up top to measure from one of these to the next.

CPU 바운드

메인 스레드 바운드 프로젝트에서 캡쳐

CPU-bound

If your game is not within the CPU frame budget, the next step is to investigate what part of the CPU is the bottleneck – in other words, which thread is the most busy. The point of profiling is to identify bottlenecks as targets for optimization; if you rely on guesswork, you can end up optimizing parts of the game that are not bottlenecks, resulting in little or no improvement on overall performance. Some “optimizations” might even worsen your game’s overall performance.

It’s rare for the entire CPU workload to be the bottleneck. Modern CPUs have a number of different cores, capable of performing work independently and simultaneously. Different threads can run on each CPU core. A full Unity application uses a range of threads for different purposes, but the threads that are the most common ones for finding performance issues are:

  • The main thread: This is where all of the game logic/scripts perform their work by default and where the majority of the time is spent for features and systems such as physics, animation, UI, and rendering. 
  • The render thread: During the rendering process, the main thread examines the scene and performs Camera culling, depth sorting, and draw call batching, resulting in a list of things to render. This list is passed to the render thread, which translates it from Unity’s internal platform-agnostic representation to the specific graphics API calls required to instruct the GPU on a particular platform.
  • The Job worker threads: Developers can make use of the C# Job System to schedule certain kinds of work to run on worker threads, which reduces the workload on the main thread. Some of Unity’s systems and features also make use of the job system, such as physics, animation, and rendering.

Main thread

The image above shows how things might look in a project that is bound by the main thread. This project is running on a Meta Quest 2, which normally targets frame budgets of 13.88 ms (72 fps) or even 8.33 ms (120 fps), because high frame rates are important to avoid motion sickness in VR devices. However, even if this game was targeting 30 fps, it’s clear that this project is in trouble.

Although the render thread and worker threads look similar to the example which is within frame budget, the main thread is clearly busy with work during the whole frame. Even accounting for the small amount of Profiler overhead at the end of the frame, the main thread is busy for over 45 ms, meaning that this project achieves frame rates of less than 22 fps. There is no marker that shows the main thread idly waiting for VSync; it’s busy for the whole frame.

The next stage of investigation is to identify the parts of the frame that take the longest time and to understand why this is so. On this frame, PostLateUpdate.FinishFrameRendering takes 16.23 ms, more than the entire frame budget. Closer inspection reveals that there are five instances of a marker called Inl_RenderCameraStack, Indicating that there are five cameras active and rendering the scene. Since every camera in Unity invokes the whole render pipeline, including culling, sorting, and batching, the highest-priority task for this project is reducing the number of active cameras, ideally to just one.  

BehaviourUpdate, the marker which encompasses all MonoBehaviour Update() methods, takes 7.27 ms, and the magenta sections of the timeline indicate where scripts allocate managed heap memory. Switching to the Hierarchy view and filtering by typing GC.Alloc in the search bar shows that allocating this memory takes about 0.33 ms in this frame. However, that is an inaccurate measurement of the impact the memory allocations have on your CPU performance. 

GC.Alloc markers are not actually timed by measuring the time from a Begin to an End point. To keep their overhead small, they are recorded as just their Begin time stamp, plus the size of their allocation. The Profiler ascribes a minimal amount of time to them to make sure they are visible. The actual allocation can take longer, especially if a new range of memory needs to be requested from the system. To see the impact more clearly, place Profiler markers around the code that does the allocation, and in deep profiling, the gaps between the magenta-colored GC.Alloc samples in the Timeline view provide some indication of how long they might have taken. 

Additionally, allocating new memory can have negative effects on performance that are harder to measure and attribute to them directly: 

  • Requesting new memory from the system may affect the power budget on a mobile device, which could lead to the system slowing down the CPU or GPU.
  • The new memory likely needs to get loaded into the CPU’s L1 Cache and thereby pushes out existing Cache lines.
  • Incremental or Synchronous Garbage Collection may be triggered directly or with a delay as the existing free space in Managed Memory is eventually exceeded.

At the start of the frame, four instances of Physics.FixedUpdate add up to 4.57 ms. Later on, LateBehaviourUpdate (calls to MonoBehaviour.LateUpdate()) take 4 ms, and Animators account for about 1 ms. 

To ensure this project hits its desired frame budget and rate, all of these main thread issues need to be investigated to find suitable optimizations. The biggest performance gains will be made by optimizing the things that take the longest time.

The following areas are often fruitful places to look for optimizing in projects that are main thread bound:

  • Physics
  • MonoBehaviour script updates
  • Garbage allocation and/or collection
  • Camera culling and rendering
  • Poor draw call batching
  • UI updates, layouts and rebuilds
  • Animation

Depending on the issue you want to investigate, other tools can also be helpful:

  • For MonoBehaviour scripts that take a long time but don’t show you exactly why that’s the case, add Profiler Markers to the code or try deep profiling to see the full call stack.
  • For scripts that allocate managed memory, enable Allocation Call Stacks to see exactly where the allocations come from. Alternatively, enable Deep Profiling or use Project Auditor, which shows code issues filtered by memory, so you can identify all lines of code which result in managed allocations.
  • Use the Frame Debugger to investigate the causes of poor draw call batching.

For comprehensive tips on optimizing your game, download these Unity expert guides for free:

CPU 바운드: 렌더 스레드

렌더 스레드 바운드 시나리오

CPU-bound: Render thread

The above screenshot is of a project that’s bound by its render thread. This is a console game with an isometric viewpoint and a target frame budget of 33.33 ms.

The Profiler capture shows that before rendering can begin on the current frame, the main thread waits for the render thread, as indicated by the Gfx.WaitForPresentOnGfxThread marker. The render thread is still submitting draw call commands from the previous frame and isn’t ready to accept new draw calls from the main thread; the render thread is spending time in Camera.Render. 

You can tell the difference between markers relating to the current frame and markers from other frames, because the latter appear darker. You can also see that once the main thread is able to continue and start issuing draw calls for the render thread to process, the render thread takes over 100 ms to process the current frame, which also creates a bottleneck during the next frame.

Further investigation showed that this game had a complex rendering setup, involving nine different cameras and many extra passes caused by replacement shaders. The game was also rendering over 130 point lights using a forward rendering path, which can add multiple additional transparent draw calls for each light. In total, these issues combined to create over 3000 draw calls per frame.

The following are common causes to investigate for projects that are render thread-bound:

  • Poor draw call batching, particularly on older graphics APIs such as OpenGL or DirectX 11
  • Too many cameras. Unless you’re making a split-screen multiplayer game, the chances are that you should only ever have one active Camera.
  • Poor culling, resulting in too many things being drawn. Investigate your Camera’s frustum dimensions and cull layer masks. Consider enabling Occlusion Culling. Perhaps even create your own simple occlusion culling system based on what you know about how objects are laid out in your world. Look at how many shadow-casting objects there are in the scene – shadow culling happens in a separate pass to “regular” culling.

The Rendering Profiler module shows an overview of the number of draw call batches and SetPass calls every frame. The best tool for investigating which draw call batches your render thread is issuing to the GPU is the Frame Debugger.

CPU 바운드: 워커 스레드

시뮬레이션성이 강한 워커 스레드 바운드의 DOTS 기반 프로젝트.

CPU-bound: Worker threads

Projects bound by CPU threads other than the main or render threads are not that common. However, it can arise if your project uses the Data-Oriented Technology Stack (DOTS), especially if work is moved off the main thread into worker threads using the C# Job System.

The capture seen above is from Play mode in the Editor, showing a DOTS project running a particle fluid simulation on the CPU.

It looks like a success at first glance. The worker threads are packed tightly with Burst-compiled jobs, indicating a large amount of work has been moved off the main thread. Usually, this is a sound decision. 

However, in this case, the frame time of 48.14 ms and the gray WaitForJobGroupID marker of 35.57 ms on the main thread, are signs that all is not well. WaitForJobGroupID indicates the main thread has scheduled jobs to run asynchronously on worker threads, but it needs the results of those jobs before the worker threads have finished running them. The blue Profiler markers beneath WaitForJobGroupID show the main thread running jobs while it waits, in an attempt to ensure the jobs finish sooner.  

Although the jobs are Burst-compiled, they are still doing a lot of work. Perhaps the spatial query structure used by this project to quickly find which particles are close to each other should be optimized or swapped for a more efficient structure. Or, the spatial query jobs can be scheduled for the end of the frame rather than the start, with the results not required until the start of the next frame. Perhaps this project is trying to simulate too many particles. Further analysis of the jobs’ code is required to find the solution, so adding finer-grained Profiler markers can help identify their slowest parts.

The jobs in your project might not be as parallelized as in this example. Perhaps you just have one long job running in a single worker thread. This is fine, so long as the time between the job being scheduled and the time it needs to be completed is long enough for the job to run. If it isn’t, you will see the main thread stall as it waits for the job to complete, as in the screenshot above. 

Common causes of sync points and worker thread bottlenecks include: 

  • Jobs not being compiled by the Burst compiler
  • Long-running jobs on a single worker thread instead of being parallelized across multiple worker threads
  • Insufficient time between the point in the frame when a job is scheduled and the point when the result is required
  • Multiple “sync points” in a frame, which require all jobs to complete immediately

You can use the Flow Events feature in the Timeline view of the CPU Usage Profiler module to investigate when jobs are scheduled and when their results are expected by the main thread. For more information about writing efficient DOTS code, see the DOTS Best Practices guide.

GPU 바운드

GPU 바운드 모바일 게임 캡처

GPU-bound

Your application is GPU-bound if the main thread spends a lot of time in Profiler markers such as Gfx.WaitForPresentOnGfxThread, and your render thread simultaneously displays markers such as Gfx.PresentFrame or <GraphicsAPIName>.WaitForLastPresent. 

The following capture was taken on a Samsung Galaxy S7, using the Vulkan graphics API. Although some of the time spent in Gfx.PresentFrame in this example might be related to waiting for VSync, the extreme length of this Profiler marker indicates the majority of this time is spent waiting for the GPU to finish rendering the previous frame.

In this game, certain gameplay events triggered the use of a shader that tripled the number of draw calls rendered by the GPU. Common issues to investigate when profiling GPU performance include:

  • Expensive full-screen post-processing effects, including common culprits like Ambient Occlusion and Bloom
  • Expensive fragment shaders caused by:
    • Branching logic
    • Using full float precision rather than half precision
    • Excessive use of registers which affect the wavefront occupancy of GPUs
  • Overdraw in the Transparent render queue caused by inefficient UI, particle systems, or post-processing effects
  • Excessively high screen resolutions, such as those found in 4K displays or Retina displays on mobile devices
  • Micro triangles caused by dense mesh geometry or a lack of LODs, which is a particular problem on mobile GPUs but can affect PC and console GPUs as well
  • Cache misses and wasted GPU memory bandwidth caused by uncompressed textures, or high-resolution textures without mipmaps
  • Geometry or Tesselation shaders, which may be run multiple times per frame if dynamic shadows are enabled

If your application appears to be GPU-bound, you can use the Frame Debugger as a quick way to understand the draw call batches that are being sent to the GPU. However, this tool can’t present any specific GPU timing information, only how the overall scene is constructed.

The best way to investigate the cause of GPU bottlenecks is to examine a GPU capture from a suitable GPU profiler. Which tool you use depends on the target hardware and the chosen graphics API.

Want to learn more?

Download the e-book, Ultimate guide to profiling Unity games, for free to get all the tips and best practices.

Was this content helpful?

Unity에서는 최적의 웹사이트 경험을 제공하기 위해 쿠키를 사용합니다. 자세한 내용은 쿠키 정책 페이지를 참조하세요.

확인