Several years have passed since the previous book, Unity Performance Tuning Bible, and Unity, devices, and the way tuning is carried out have all changed. This prologue first gives a broad review of the changes in the engine and devices, and then describes where tuning is heading as of 2026.
The biggest change in Unity is in rendering. With Render Graph introduced to URP in Unity 6, the focus of optimization work has shifted from rearranging commands yourself to correctly declaring resource usage to the engine and verifying the result. The Unity 6 generation also adds new options for shader warmup and for passing per-object parameters. On the C# side, little changes within Unity 6, but Unity 7 introduces CoreCLR to the Editor and some desktop players, and the garbage collection mechanism also changes significantly. Mobile, on the other hand, will stay on IL2CPP and the Boehm GC for the time being, so the existing approach to memory optimization remains valid. Around build and distribution, there are more settings that let you adjust the balance of size, memory, and speed without rewriting code.
Devices have also changed greatly. Comparing GPU Geekbench scores*1, the iPhone 14 Pro (A16), a high-end device in 2022 when the previous book was written, scored 18,691, while the A20 Pro in the 2026 iPhone 18 Pro and iPhone Duo scores 55,120, about three times higher. On the Android side as well, the Galaxy S22 (Snapdragon 8 Gen 1, Adreno 730) scores 6,179, while the Galaxy S26 Ultra (Snapdragon 8 Elite Gen 5 for Galaxy, Adreno 840) scores 27,501, more than four times higher.
Low-end devices, however, have not changed much. The Galaxy A23 5G, released in Japan in 2022, had a Dimensity 700 (Mali-G57 MC2) and 4 GB of memory. The Galaxy A25 5G of the same series, released in 2025, uses a different SoC, the Dimensity 6100+, but has the same Mali-G57 MC2 GPU with a score of 1,400 and still 4 GB of memory. This is a gap of about 20 times compared with the Galaxy S26 Ultra.
On iPhone as well, in 2022 devices with 3 GB of memory or less, including the iPhone SE (2nd generation), were still widely used. In 2026, devices with 6 GB of memory or more, from the iPhone 14 onward, have become mainstream. However, devices with performance sufficient for everyday use were already widespread by 2022, so people are less inclined to replace their devices even when new ones are released. Many 4 GB devices, such as the iPhone SE (3rd generation) and iPhone 13, which were sold until early 2025, remain in use today, and the 3 GB iPhone SE (2nd generation) is still used in some cases.
As a result, the question of which devices to target, which has long troubled Android, can no longer be ignored on iOS either. The range of devices a single title must support has widened, from the iPhone 18 Pro with 12 GB of memory at the top to the iPhone SE (2nd generation, released in 2020) with 3 GB at the bottom. Depending on the game genre, switching rendering expression and quality settings according to device performance is required more than ever (Part 3, Chapter 3, "Switching Rendering Settings Per Device").
[*1] Geekbench Browser https://browser.geekbench.com/
Even so, the problems we are trying to solve have not changed much since the previous book: mobile GPU memory bandwidth, CPU cost when drawing many objects, spikes from shader compilation, leaked asset releases, and per-device memory limits. Most of the new features are new tools for these existing problems. For this reason, this book does not cover only features available in Unity 6 or later. It also includes much content that can be applied as is to projects on earlier versions, such as countermeasures for character animation cost, approximation of transcendental functions in shaders, lifetime management of AssetBundles, and raising the iOS memory limit.
At the beginning of Parts 1 to 3, we again summarize what has changed in that area in 2026 and what each chapter covers. If you want to find a chapter from a symptom, refer to the "Problem-Based Reverse Lookup Guide" at the front of the book.
On the development side, the biggest concern in 2026 is AI. AI coding agents are now used daily as tools to delegate part of implementation and review, and performance tuning is no exception. Part 4, Chapter 3, "Turning the Tuning Bible into a Skill for AI Coding Agents", covers an attempt to convert the knowledge of the previous book and this book into evaluation rules for agents.
However, tuning has a prerequisite that comes before delegating to AI: measurement. If you make changes without measuring where the slowness is, you cannot judge whether things got faster. For this reason, the industry as a whole is first working on automating measurement. The direction is to collect physical-device numbers with a fixed procedure for every build and automatically detect regressions by comparing them with past measurements, which reduces how often people open the Profiler by hand and prevents late discovery.
At CyberAgent's Game & Entertainment division, to which the authors belong, we are also building, as part of our automation work, a runtime bridge that lets AI agents observe and operate Unity apps in real time. Combined with a device farm, this setup lets agents handle everything from CI, from reserving a device, installing the app, operating it with a defined procedure, to measuring and pulling back the results. For measurement, we use a runtime profiler embedded in the app on a physical device, send the collected data to a server, and make it possible to compare it with past measurements in a browser and to be notified of changes that exceed a threshold. An AI agent can later retrieve the data sent to the server and carry out everything from diagnosing bottlenecks to fixing them (Part 4, Chapter 2, "Automating Physical-Device Profiling of Daily Builds").
Of course, this does not mean human judgment is no longer needed. People still need to hold the criteria for what to look at, what to compare against, and where to stop. Only the parts where those criteria already exist can be automated. This is also why this book presents material for judgment together with measured values.