Tinker Diffusion language models

by Samy (Kondasamy Jayaraman) · drawn with Colorfigures by April Zhu

1 token a pass × 150 passes a second = 150 tokens a second; a 16-token reply in 16 passes.

Chapter 6 of 7

Reading a speed claim

The vendor’s figure is on the sales deck. Larkspur’s team wants to know what sits behind it before signing.

This chapter: What makes up tokens a second, and what should you ask about it?

About this chapter

You will learn

  • How tokens a second splits into two numbers
  • What a speed claim leaves out
  • What to ask a vendor

Words you’ll meet

Tokens a second
Tokens a pass times passes a second. A vendor can raise either one.
Passes a second
How many times a second the hardware runs the model over the reply. Here, Larkspur’s planning number.
Vendor claim
A figure the seller measured on its own setup and reports itself.

Step 1 of 3

A : 1,009 . Larkspur’s server runs 150 , its planning number.

Predict first.

What does a model’s tokens a second depend on?