Getting started

Run a test

One click, about 50 seconds, five phases. What happens, and how to get a clean number.

A full run, to scale. Gaps are 650 ms transitions where the GPU idles and the visuals run at full frame rate.

Requirements

Chrome, Edge	113 or later on Windows, macOS and ChromeOS. On Linux, enable chrome://flags/#enable-unsafe-webgpu.
Safari	26 or later on macOS. The GPU model is hidden, so runs are measured but unranked.
Firefox	141 or later on Windows. The model string is sanitized, so runs are unranked.
GPU memory	512 MiB free. The engine halves the working set if allocation fails, and marks the run reduced.
Page	HTTPS (a secure context). WebGPU is unavailable on insecure origins.

Close games, video and other GPU-heavy tabs. Shared GPU work lowers every phase.
Plug a laptop in. Battery modes often run memory at lower clocks.
Keep the tab visible. A hidden tab is recorded as a note on the result.
On a laptop, tick “This computer is a laptop” so the run can enter Laptop Week.

Phase	Budget	On the stack
Read	6.5 s	Cyan pulses fall from the dies to the base die.
Write	6.5 s	Violet pulses rise into the dies.
Copy	6.5 s	Pulses in both directions.
Random	4.5 + 3 s	Scattered flicker, then a single pulse walking the chain.
Sustain	18 s	A continuous stream, drifting from cyan to violet.

Layers light one by one as the run progresses. All sixteen are lit when the result appears.

MEM Score: Log scale. Reference 50, ten times the reference 100.
Percentile: Position inside your device class. Needs 5 verified runs.
Device class: GPU model and its published memory spec.
Read, write, copy: Median effective GB/s. Read also shows the share of published peak.
Random, latency, sustain: Scattered throughput, dependent-load latency proxy, retention under load.
AI decode ceiling: Upper bound on tokens per second: read GB/s divided by model bytes.
MEMPRINT: A signature of this run, not of your machine.
Stack position: Layer and die number in The Stack.

You see	Cause	Fix
WebGPU is unavailable	Unsupported browser, no adapter, or blocklisted driver.	Use a browser from the table above; update GPU drivers.
Could not allocate memory	Less than 512 MiB of GPU memory free.	Close other GPU apps and tabs.
The GPU stopped answering	Device lost: driver reset or another app claimed the GPU.	Close GPU-heavy apps, run again.
Unstable	Samples varied by more than 25 %.	Remove background load, plug in, run again.
Unranked, model hidden	Safari or Firefox hide the GPU model.	Run in Chrome or Edge to join a class.
Timed with the wall clock	The browser did not grant GPU timestamps.	Informational. Samples double in length to keep error small.