ContentsThe library

Training Across Many Machines

You Do Not Choose One, You Multiply Four

Last timeNobody Holds the Whole Thing

The four ways of dividing training are not alternatives. They compose, and each one is assigned to the link whose speed its traffic can live with.

Five lessons have each presented one way to divide the work, and each one ended

with a limit. That is not four candidates to pick between. It is four axes, and

a real training run uses all of them at once.

FIG 1What each split puts on the wire
messages per stepmegabytes per messagecollective, not point tolink speed it needs, GB practical degree
same model, different da1200251024
same layer, split across16010014008
different layers, differ63002516
state split, weights gat360140064
The two marked cells are the design. Splitting a layer needs four hundred gigabytes a second and will not work without it; splitting by stages is content with twenty-five, which is an ordinary network. Everything else in this lesson follows from those two numbers being a factor of sixteen apart.

The four axes

Read the rows as traffic signatures rather than as techniques. The data split

sends one large message at the end of a step, which is why a larger per-device

batch helps it. The inside-layer split sends a hundred and sixty medium

messages that nothing can hide, because the next operation in the layer is

waiting on each one. The stage split sends a handful of messages to one

neighbour each, which can be overlapped. The state split adds gathers that

scale with the model rather than the batch.

They multiply

Each axis has a practical ceiling, and all four ceilings are low. Beyond eight

or sixteen, splitting inside a layer runs out of fast cabling. Beyond sixteen

stages, the idle fraction demands a batch nobody can afford. The data split

runs out when the per-device batch falls too low to keep a device busy.

The resolution is that the degrees multiply.

The lesson stops here

4 more paragraphs to go

You have read the opening. The rest of the argument, the problems that check whether it landed, and the lines worth keeping at the end all come with a plan.

The first lesson of every course in the library reads the whole way through, free, so you can see exactly what the rest of them are.

See the planThe contents

This is the reading half

Starting the course gives you your own copy of it. Every idea on every page has problems standing under it, marked with a reason rather than a tick, and any sentence you do not believe can be opened and argued with. None of that can happen on a page nobody owns.

The contents

The rest of this course

  1. 01The Weights Are the Smallest Thing You Have to Store
  2. 02Every Machine Does the Whole Thing, and Then They Agreeopening only
  3. 03Nobody Is in Charge and That Is Why It Scalesopening only
  4. 04Half a Layer Each, and a Phone Call in the Middle of Every Oneopening only
  5. 05Four Machines, and Three of Them Are Waitingopening only
  6. 06Sixteen Copies of the Same Numbers, in Sixteen Placesopening only
  7. 07You Do Not Choose One, You Multiply Fouryou are here
  8. 08Three Hundred Devices Waiting on Oneopening only

Read alongside