xavier-ramirez.com

GPU or CPU?

You built a network out of dot products. Now watch it run. The same math — a mountain of multiply-and-adds — is why AI is built on GPUs, wired together by the thousands.

PROCESSING A NEURAL NETWORK
layer by layer — the same network, two machines
CPU· one neuron at a time
GPU· a whole layer at once

Every neuron in a layer is independent — so the GPU computes them all at once, while the CPU grinds through them one by one.

Remember the neuron from the last lesson: an output is w · x — multiply each weight by each input, add them up. A layer is just many neurons doing that side by side. To run the network you compute layer 1, feed it into layer 2, then layer 3 — a stack of these steps.

Here is the key fact: within a layer, every neuron is independent — none needs another's answer, so they can all be computed at the same time. Only the layers depend on each other, in order.

A CPU has a handful of large, clever cores; it mostly computes one neuron, then the next, then the next. A GPU has thousands of small cores, so it computes a whole layer's neurons at once. Same math — wildly different speed.

TRY IT · RUN THE NETWORK
Watch it process, layer by layer

The same network runs on both machines. Press run and watch: the GPU lights up a whole layer at a time; the CPU crawls neuron by neuron. Add neurons and the gap explodes.

Each neuron computes one dot product:σ(w·a) = σ( Σᵢ wᵢaᵢ )
CPUa few cores
σ(w·a)· 1 at a time
ready
GPUthousands of cores
σ(w·a)· 6 cores at once
ready

each neuron = one w · x dot product over the layer before it

neurons
15
parameters
93
GPU finishes sooner by

The GPU computes an entire layer in one shot, because its neurons don't depend on each other. It still waits for each layer before the next — layer 2 needs layer 1's answers. The CPU gets no such shortcut: it walks every neuron, one by one. Widen the layers and its bar runs off the edge.

This toy has a few dozen neurons. A real model has billions of parameters and hundreds of layers — which is why a CPU would take months on what a GPU cluster does in days.

So a GPU is the right tool. But one GPU is not enough. Each has a fixed pool of fast memory — say 80 GB — and a frontier model's weights are hundreds of gigabytes. The model simply doesn't fit.

So we wire many GPUs together. Eight GPUs share a board as a node, linked by ultra-fast NVLink. Many nodes fill a rack, linked by InfiniBand networking. The GPUs are constantly swapping half-finished results — so how fast those wires are matters as much as the GPUs themselves.

TRY IT · INSIDE A RACK
The wires between the GPUs

GPUs on the same node talk over NVLink — blazing fast. GPUs on different nodes cross the InfiniBand network — far slower. Send a tensor and watch where it has to travel.

Send a tensor between two GPUs:
LINK
NVLink
BANDWIDTH
900 GB/s
RELATIVE TIME
HOPS
1

The lesson every distributed-training engineer learns: keep the chatty work inside a node. Every hop onto the network costs you — so you arrange the model to cross it as little as possible.

So: does each GPU hold the whole network, or just a piece of it? Both happen — it depends on the strategy. Here's the same next-word network on 2 GPUs, three different ways.

TRY IT · SPLIT THE MODEL
One network, spread across GPUs

This little network reads a prompt and predicts the next word. Pick a strategy and watch what each GPU actually holds — the solid neurons live on that GPU; the faint ones live on the other.

one network · predicts the next word
Each GPU holds:the whole network — a full copy

Every GPU holds a complete copy of the network and runs a different batch of prompts. After each step they average their weight updates so the copies stay identical.

GPUs must exchange:share gradients (all-reduce)

Real training combines all three at once — data and tensor and pipeline parallel — across thousands of GPUs for weeks. A frontier model is far too big to copy whole onto one GPU, so it's both split up and replicated. That's what a training run really is, and why it costs what it costs.

EXPLAIN IT BACK
Why is a matrix multiply the perfect job for a GPU, and a bad fit for a CPU?
NEXT: EMBEDDINGS

You've followed a network from math to metal. Next: the very first thing that math touches — how a token becomes a vector of meaning the network can compute on.

Neural networksEmbeddings
Language: English