Quiz · 5 questions
📐 Linear Algebra for AI
Why every model collapses to dot product + GEMM — and how the roofline keeps you honest
Level 0Beginner
0 XP0/38 lessons0/12 achievements
0/100 XP to next level100 XP to go0% complete
Quiz
01Which BLAS level has the best GPU suitability and why?
Hint
Think about how many times each loaded byte is reused.
02What does 'coalesced access' mean on a GPU?
Hint
It's about the access pattern of neighboring threads.
03A kernel with low FLOPs per byte (< 4) is typically:
Hint
Few FLOPs per byte loaded = lots of waiting on memory.
04Why does cuBLAS default to column-major storage?
Hint
It's a historical artifact, not a hardware requirement.
05Why does batched transformer inference run far faster per token than single-token decoding?
Hint
Per-token, the FLOP count is the same. What changes is the intensity.
Comments 0
🔔 Reply notifications (sign in)Sign in — Please sign in to comment.
No comments yet — be the first.