How a Browser Trains a Neural Network: Tensors as Textures, Ops as Shaders

You can open a tab, click a button, and watch an LSTM's loss curve go down. No server, no Python environment, no CUDA. How does that work? The short answer: the browser borrows your graphics card through WebGL, and a few clever mappings turn a graphics API into a math engine. Why a GPU at all? Training a neural network is mostly matrix multiplication: the same simple operation repeated over thousands of numbers. A CPU does a few things at a time very fast. A GPU does thousands of simple things at the same time. That is a good match. The catch is that WebGL was designed for drawing triangles, not for multiplying matrices. The trick: tensors become textures, ops become shaders Libraries like TensorFlow.js use two mappings: Machine learning WebGL Tensor (an array of numbers) Texture (an image stored in GPU memory) Operation (matmul, add, ReLU) Fragment shader (a small program that runs once per pixel) Think of a texture as a grid of pixels, where each pixel's color channels hold numbers instead of colors. To "compute" something, the library draws a rectangle that covers the output texture. The GPU runs the fragment shader once for every output pixel, in parallel, and each run computes one element of the result. JavaScript -> WebGL API -> GPU driver -> graphics hardware Nothing is being drawn on screen. The "image" is the answer. What one training step looks like Forward pass. Input tensors are uploaded as textures. Each layer is a chain of shader programs that read one texture and write another. Loss. A final shader reduces the prediction error to a number. Backward pass. Gradients are just more operations, so they also run as shaders. Update. The optimizer (SGD, Adam) updates the weight textures, again on the GPU. The important detail: data stays on the GPU between steps. Only small things, like the loss value you want to print, come back to JavaScript. A minimal example import * as tf from '@tensorflow/tfjs'; await tf.setBackend('webgl'); await tf.ready(); const model = tf.sequential(); model.add(tf.layers.lstm({ units: 8, inputShape: [5, 1] })); model.add(tf.layers.dense({ units: 1 })); model.compile({ optimizer: tf.train.adam(0.01), loss: 'meanSquaredError' }); await model.fit(xs, ys, { epochs: 5, batchSize: 32, callbacks: { onEpochEnd: (epoch, logs) => console.log(epoch, logs.loss), }, }); Here xs and ys are tensors of shape [samples, 5, 1] and [samples, 1]. Everything between fit and the callback runs through the mappings above. Where it hurts WebGL gets the job done, but it has real limits. Reading data back is slow. Calling something like dataSync() forces the CPU to wait for the GPU. Do it once per epoch, not once per step. First run is slower. Shaders have to be compiled the first time each operation is used. Memory is not garbage collected. Textures live on the GPU. Wrap temporary work in tf.tidy() and call dispose() on tensors you no longer need, or memory will creep up. Texture size and precision limits. Large tensors are split across textures, and some devices, especially phones, may use lower precision than you expect. No general-purpose compute. WebGL has no compute shaders, so every operation is squeezed into a rendering pass. Newer APIs such as WebGPU are designed to fix this, but browser support still varies, so check before relying on it. Background tabs get throttled. Browsers slow down timers and pause animation frames in hidden tabs, so a training loop can stall when you switch away. Keeping the tab in the foreground is the simple fix. Practically, this means small to medium models: LSTMs with a few hundred units, small Transformers, time-series experiments. Not billion-parameter models. Why bother training in the browser? Privacy. The data never leaves the user's device. Zero server cost. The user's GPU does the work. Instant demos. A link is all anyone needs. The trade-off is scale. For learning, prototyping, and visualizing how training behaves, it is hard to beat. Try it yourself I built a small tool around this idea: you can train LSTM and Transformer sequence models on time-series data entirely in your browser, with live loss curves, at ai-spinach.xyz. The write-up above is part of its technical docs. If you have hit different limits with WebGL training, or you are experimenting with WebGPU, I would like to hear about it in the comments.