GPU Benchmark

How it works

1. Close Other GPU-Heavy Tabs and Apps. On a laptop, plug in and set the power mode to performance first — battery-saving modes cap GPU clocks invisibly, the same way they cap frame rate on the FPS test.
2. Read the Capability Report. It appears the moment the page loads, needs no permission, and runs no GPU load at all — just what your browser is willing to tell a web page about your graphics hardware.
3. Confirm the Safety Notice. The benchmark deliberately loads your GPU hard: that uses more battery and produces some heat, and the page will feel slow while it runs. Read it once, then confirm.
4. Run the Quick Benchmark. About 8 seconds. Read the three throughput figures — shader fill rate, draw call throughput, texture sampling rate — each with its own unit and uncertainty.
5. Run the Sustained-Load Check If You're Chasing Heat. A separate 30-second button. If your machine gets slower the longer you use it, this is the check that actually tests for it, by comparing early throughput against late throughput.
6. Read the Summary, Not Just the Numbers. Hardware acceleration, GPU selection, and context stability are the findings that matter most. The raw throughput rates are context for comparing your own machine over time, not a score to compare against someone else's.

What This Test Actually Measures

Every stage of the benchmark renders to an off-screen surface that is never shown on your monitor and never composited — so nothing about the result can be capped by your display's refresh rate the way a visible animation would be. The reported number is work completed divided by real elapsed time: pixel-shader iterations per second, draw calls per second, or texture samples per second. That is a genuine measure of GPU throughput, and it is a structurally different thing from a frame rate.

Reading Your Capability Report

WebGL 2 is the modern baseline nearly every current browser supports; WebGL 1 only is a real but shrinking minority, mostly older devices; no WebGL at all means this page can't measure or report anything further about your graphics hardware.

The limits table below it is a direct readout of what your graphics driver reports back to the browser — maximum texture size, how many samples it can anti-alias with, how many textures a single shader can sample at once, and similar hard ceilings. The extensions list picks out a curated handful worth explaining (floating-point render targets, anisotropic filtering, compressed texture formats) out of the 30-40+ a real browser typically supports.

Why Your GPU's Name Might Be Hidden

Safari always returns a constant string — "Apple GPU" — never a real model name, on every Mac and iPhone regardless of which chip is actually inside. Firefox generalizes or withholds the string entirely when its own fingerprinting-resistance setting is turned on. This is a deliberate browser privacy decision, not a bug on this page's part, and a constant value is not a measurement — so rather than dressing a placeholder string up as a real GPU model, this page states plainly that the name is hidden and explains why.

None of this affects the throughput numbers below it. A hidden name and a fully-measured shader fill rate are completely independent — your browser can refuse to name your GPU while this page still measures exactly how fast it is.

Software Rendering: the Single Most Important Thing This Page Can Tell You

A software renderer means your CPU is pretending to be a graphics card — every pixel is computed in software, at a small fraction of real hardware speed, whether or not the reported name looks like a real GPU. This page detects it two ways at once: a decisive software-family renderer string (SwiftShader, llvmpipe, Microsoft Basic Render Driver), or a hardware-looking name whose measured throughput is far below what any real GPU produces — the second case usually means a degraded or emulated path hiding behind a normal-sounding name.

• Hardware acceleration turned off in your browser's own settings.
• A specific GPU driver on your browser's internal blocklist, often after a known driver bug.
• Running inside a remote desktop or virtual machine session, where the real GPU usually isn't passed through at all.
• A badly outdated or corrupted graphics driver.

Sustained Load and Thermal Throttling

The 30-second sustained check repeats the shader stage continuously and compares the first 5 seconds of throughput against the last 5, reporting how much of the starting rate survived. A fanless phone or a thin ultrabook losing a real chunk of throughput after 30 seconds of full load is normal engineering, not a fault — small, sealed devices are built to throttle rather than overheat, and this page frames the result that way rather than as a verdict on your hardware.

Laptops With Two GPUs: Why the Fast One May Never Engage

A gaming or creator laptop with both an integrated and a discrete GPU normally lets an application request either one, via a hint called powerPreference. This page requests BOTH — a low-power context and a high-performance one — and compares the two adapters' identities. Two genuinely different renderer strings means the fast GPU really did engage for the high-performance request.

The same renderer string on both requests does not prove your dGPU is broken — a single-GPU machine produces exactly the same result, completely normally. It is real, useful information on a machine that is SUPPOSED to have two adapters, though: it can mean a hybrid-graphics driver setting, a MUX-switch configuration, or a genuine fault worth investigating with your laptop's own GPU-selection software before assuming the worst.

What This Test Is Not

• Not a VRAM reading — no web API exposes how much video memory your GPU has.
• Not a cross-device score — there is no composite number, no tier ladder, and no percentile ranking anywhere on this page.
• Not a temperature or wattage reading — no browser API exposes either one.
• Not a refresh-rate or FPS measurement — see the FPS test for that.
• Not a driver-level GPU timer. The browser extension that would enable one, EXT_disjoint_timer_query_webgl2, exists only in Chrome and Edge and is disabled by default even there, so using it would produce one number on Chrome and a structurally different one everywhere else — this page's timing is plain wall-clock instead, which behaves identically on every browser.

Frequently asked questions