KERNELIN PROJECTS

Real GPU workloads, written in plain Kotlin.

Three applications built with Kernelin 0.1.0 from Maven Central. Each runs the same Kotlin functions on your GPU (WebGPU) and on the CPU, checks its own answers against exact formulas or ground truth, and shows measured speed. Open one in Chrome, Edge or Safari, on desktop or Android.

Simulation · finance

Risk Engine

Monte Carlo pricing and greeks of an options book: European, Asian, barrier and lookback, over 252 trading days.

46 ms

1M paths × 4 scenarios on an Apple M5 GPU · 53× faster than 10 JVM cores

Checked on every run against Black–Scholes and the geometric-Asian closed form.

Open app Code

Image · photo editing

Photo Studio

Live photo editing: edge-preserving denoise, exposure, white balance, tone, color, sharpening, vignette.

+11.3 dB

denoise of 3.8 MP in 48 ms on the GPU · 119× faster than in-browser CPU

Measured as PSNR against a noise-free test chart with real sensor noise.

Open app Code

Video · camera analytics

Motion Guard

Traffic and security camera analytics: motion vectors for every block, vehicle speed per lane, zones and events.

0.995 / 0.972

precision / recall vs ground truth · 576 M comparisons per 720p frame in 19 ms on the GPU (53× in-browser CPU)

Lane speeds within 6% of the truth; swaying trees and noise never alarm.

Open app Code

How to read these numbers

GPU times include copying data to the GPU and back. CPU times are the same Kotlin functions on the JVM (all cores) or compiled to Wasm (one browser thread).
Every project has unit tests that run the kernels as normal Kotlin, so the logic is tested before it ever reaches a GPU.
Building them found a real bug in Kernelin (helpers named like WGSL built-ins, such as uniform); it is fixed for the next release.