Chapter 2 · Models

Chapter 2 exam

The full chapter, fifteen questions. Clear it to move on to hardware.

Builds on

Chapter progress

0 / 17 lessons completed

Up next in this chapter: GPU compute →

This is the whole chapter in one sitting. Larry pulled the engine apart — neural nets, transformer blocks, attention and the KV cache, MoE, diffusion, and the roofline that governs all of it. Fifteen questions below decide whether it stuck. Clear it and Chapter 3 opens.

Score 80% or better to pass. Every option explains itself, so a wrong pick still teaches — and you can take another pass whenever you like.

01 · The exam

Fifteen questions across the chapter.

Parameters and FLOPs, the KV cache and GQA, MoE, arithmetic intensity and the roofline, FlashAttention and PagedAttention, and the diffusion/video cost model.

Check your understanding0 / 15

Question 1

A matmul multiplies an M×K matrix by a K×N matrix. Roughly how many floating-point operations does it take?

Question 2

Per token, the KV cache stores bytes proportional to which product?

Question 3

Llama-3 70B uses 8:1 GQA (8 KV heads vs 64 query heads). Versus hypothetical MHA (64 KV heads), the KV cache shrinks by roughly:

Question 4

In a Mixture-of-Experts layer with top-k routing, what is true of the parameters used per token?

Question 5

Arithmetic (operational) intensity is defined as:

Question 6

On a roofline plot, the ridge point (where memory-bound meets compute-bound) sits at intensity equal to:

Question 7

Why is single-request LLM decode typically memory-bandwidth-bound?

Question 8

Prefill (processing the prompt) differs from decode because it:

Question 9

FlashAttention is faster mainly because it:

Question 10

Online softmax lets attention be computed tile-by-tile in one pass by tracking, per query, a running:

Question 11

PagedAttention improves serving throughput chiefly by:

Question 12

A model with a 4K sliding-window attention is asked about information at position 20K. What happens?

Question 13

Modern LLM FFNs favor GELU/SiLU over ReLU mainly because they are:

Question 14

Few-step image generation (distilled diffusion) reduces cost primarily by:

Question 15

Video generation is far costlier than image generation mainly because: