Sixty frames per second on ProMotion
Design shipped a refresh of our 80-card catalog grid: layered drop shadows, backdrop-filter: blur(12px) on status badges, rounded corners, and a parallax hero banner driven by a scroll listener. During design review on Karim's M3 MacBook Pro with a 120 Hz ProMotion display, scrolling was silk at 60 FPS and nobody opened Chrome's Layers panel.
Four minutes after mobile rollout, Android telemetry fell off a cliff. Frame rates dropped to 11 FPS on Moto G and Pixel 6a phones, main-thread frame times jumped past 88 milliseconds, and 14% of mobile tabs died with RESULT_CODE_HUNG or a GPU process OOM crash within thirty seconds of scrolling. Inès plugged the Moto G into her laptop over USB, opened chrome://inspect, and watched the renderer memory graph climb past 380 MB before the tab turned into a gray sad-face icon.
Doubling the crash rate with hardware acceleration
Karim figured the phone was falling back to software paint rasterization because the cards lacked GPU promotion. He added two lines to .product-card so every card would get its own hardware layer:
.product-card {
will-change: transform;
transform: translateZ(0);
}
On his 16 GB unified-memory MacBook, it stayed at 60 FPS. On Inès's Moto G, the tab crash rate doubled.
Promoting all 80 cards at once forced Blink to allocate 80 separate GPU compositor layers in VRAM on initial load. Each card's backing store, multi-stop shadow buffer, and backdrop-filter readback surface consumed roughly 4.5 MB of GPU memory. On a desktop with 16 GB of unified memory, 380 MB of card textures is a rounding error. On a mid-range Android phone sharing memory between the OS, the renderer, and the GPU, allocating 80 permanent textures before the user's thumb even touches the glass pushes the GPU process straight into an OOM kill. Making every card its own layer also meant the compositor had to upload 80 textures across the bus before the first scroll could even settle.
Why the main thread was still sweating
Even before will-change, our parallax hero banner was running a window.addEventListener('scroll', ...) handler that read window.scrollY and mutated inline transform styles on every scroll tick. Combined with backdrop-filter: blur(12px) sitting over scrolling cards, Blink had to re-rasterize blurred badges and multi-stop shadows on the CPU every frame, keeping the main thread pinned above 88 ms per frame.
Julie stripped will-change: transform off .product-card, replaced backdrop-filter: blur(12px) on the 80 repeating badges with a crisp semi-opaque background, and replaced the JS parallax listener with a CSS scroll-driven animation (animation-timeline: scroll()), moving the parallax transform onto the compositor thread with zero JavaScript per frame.
Containment and skipping offscreen cards
The postage-stamp scrollbar trap
First try with content-visibility: auto: the scrollbar thumb shrank to the size of a postage stamp, and every time Inès scrolled down a screen, the scrollbar jumped backward and the page height bounced.
.product-card {
contain: layout paint;
content-visibility: auto;
contain-intrinsic-size: auto 320px;
}
Without contain-intrinsic-size, Blink treats every skipped offscreen card as 0px tall. The document initially thinks it is only as tall as the first visible row, then inflates by 320 px per card as you scroll down, pulling the scroll position out from under your thumb. Adding contain-intrinsic-size: auto 320px gave Blink a stable placeholder height and remembered each card's real measured height once rendered.
What stayed in the codebase
A modern laptop GPU forgives sins that kill a 200-euro phone in four seconds.
Our team rule is simple now: nobody is allowed to say will-change out loud without scrolling the page on the phone from Inès's drawer first. Tab crashes dropped from 14% to 0.02% within an hour of the fix going live, and the Moto G went back to sleeping peacefully in her drawer until the next CSS experiment.

Top comments (0)