ApexMark: A Benchmark That Runs in the Browser

Comparing machines is harder than it looks. Native benchmarks need an installation and do not exist everywhere, and the benchmarks that already live in the browser mostly measure the browser’s JavaScript engine rather than the device underneath. A score is only worth trusting if it cannot be flattered: by a huge processor cache that hides the speed of the memory, by a scene that happens to favour one graphics vendor, or by a result that quietly skipped a test. The challenge was to build a benchmark that anyone can open from a link, that pushes a machine hard enough to mean something, and that earns the right to be believed.

The product is a measurement instrument, so the design work was mostly about honesty. Five disciplines run in sequence and every workload is visible: the processor ray-traces an image row by row across all its threads, the graphics card climbs a 15-level WebGPU ladder from 360p to 4K while holding 30 frames per second and is then put through validated compute (arithmetic, matrix multiplication, buffer bandwidth), the memory test draws the real addresses it visits, the storage test paints the blocks it writes and reads, and the internet test plots the bytes as they arrive from a Cloudflare edge. Nothing counts until its output has been checked, and a test the browser cannot run is reported as unavailable, never as a zero. The overall index is a geometric mean of CPU, GPU and memory (7 : 7 : 3, where 1,000 is the frozen ApexMark 1.0 baseline), chosen after studying how PassMark, 3DMark, AnTuTu, Geekbench and PCMark combine their scores. Storage and internet are shown but kept out of the index on purpose: browser storage is served through caches, so it is not a raw SSD figure. The most instructive moment came from real data. Comparing two leaderboard machines showed a memory latency almost five times lower than on a comparable machine: the test’s 64 MiB table fit entirely inside a 96 MB L3 cache, so it was measuring the cache, not the memory. The memory test was rebuilt as a ladder that grows beyond any processor cache and scores the plateau, and the original engine was kept frozen beside it. Engines are versioned, their protocols are pinned, and their scores are never ranked together, so history stays honest. Graphics run through a single portable shader path with no vendor-specific code, and cross-vendor comparison is part of the validation.

ApexMark is live as an experimental benchmark: run it as a guest in about two to three minutes, export the evidence, or sign in to keep a private history and publish a score to the Hall of Fame. The result is an instrument that can be pointed at almost any modern device from a link, and whose credibility comes from what it refuses to claim.