How Google Optimized Chrome to Set New Records in Benchmarks
The Google Chrome team has published an in-depth technical report on the optimizations introduced in the browser over the past year, documenting the progress made on key industry benchmarks.
Measuring was conducted on a MacBook Pro with an Apple M5 chip running macOS 26.0.1 and shows results that the team describes as absolute records for any currently available browser: 469 points on JetStream 3, with a 10% increase since early 2025, and 61 points on Speedometer 3.1, up 5% from the previous update reported about a year ago.
The area where the team focused most of its work is the V8 JavaScript engine. A significant part of the improvements involved the introduction of so-called “fast paths” in the optimizing compiler: instead of executing the entire standard implementation for every integrated function, the compiler identifies the most common cases and handles them through shortened execution paths, eliminating unnecessary steps.
On the optimization heuristics front, the team revised the criteria with which V8 decides which functions to optimize and when to do so. Optimizing a JavaScript function incurs a computational cost that must be justified by the expected gain in repeated execution. In this context, the behavior of asynchronous functions and generator functions has been improved, allowing the engine to optimize them early when executed in parallel in large numbers. The criteria for inlining between different functions have also been refined, taking into account the types of objects exchanged between them in order to reduce internal allocations in the engine.
Another area of intervention focused on managing BigInt values, a category of objects that emerged as a bottleneck in analyses conducted on the new JetStream 3. The algorithms for division and canonicalization have been improved, and concurrently, the data allocation structure related to these objects was modified, with a stated benefit also in terms of security within the browser sandbox.
Work on WebAssembly support has built upon progress already achieved in previous versions, which included eliminating redundant loads and null-checks, as well as supporting speculative inlining of indirect function calls with a deoptimization mechanism. Based on this, the analysis of new entries introduced in JetStream 3 made it possible to identify further margins of improvement.
Among the most significant interventions are the optimization of code generation for SIMD (Single Instruction, Multiple Data) instructions and the improvement of register allocation, with documented positive effects on workloads related to artificial intelligence, cryptography, and interpreter execution. Memory reuse within the compiler has also been improved, benefiting compilation speed in certain scenarios.
A second line of work addressed the overhead of JavaScript code calls to WebAssembly modules. The compiler can now eliminate repeated type conversions and redundant loading operations within loops, integrating the call management code with existing optimizations and allowing V8 to more effectively remove unnecessary conversions.
The third area of intervention concerned Blink, the Chrome rendering engine. To reduce style resolution times and DOM operations, a unified cache for the querySelector() method was introduced, aiming to eliminate redundant lookups. Management of HTML element attributes was optimized through a fast exit path that avoids unnecessary checks when attributes are absent. The caching mechanism for dataURI resources was also extended beyond the scope of a single document, benefiting modern web frameworks that extensively use this type of resource. Style recalculation for CSS transitions has been streamlined and the CSS selector matching system has been simplified to enhance animation fluidity.
Regarding HTML parsing and text manipulation—operations that significantly impact initial page load times—SIMD operations have been introduced to make string copying significantly more efficient. The creation of strings directly in the HTML parser in fast-path mode has also been optimized, reducing overhead in constructing DOM elements from raw text.
On the typography and vector graphics front, the team has completed optimizations for text formatting in Apple Advanced Typography (AAT), resolved several issues regarding font fallback, and eliminated unnecessary heap allocations during glyph width calculations. To accelerate the rendering of complex SVG graphics, a dedicated cache has been introduced. At the same time, the internal data structures of DOM storage have been simplified, reducing memory usage and lightening fundamental page layouts.