3.9Multi-level cache
The idea
A multi-level cache puts a small fast L1 in front of a bigger slower L2. The average access time adds the penalties level by level.
Think of it like this: Check your pocket first (L1), then your bag (L2), then go home (memory).
Why it exists
One cache cannot be both very fast and very large; two levels give nearly the speed of the small one and the hit rate of the big one.
How it works
- Every access pays t1 (L1 lookup).
- Only on an L1 miss do we also pay t2.
- Only if L2 also misses do we pay memory time.
- AMAT = t1 + m1 × (t2 + m2 × t_mem).
Worked example: t1 = 2, h1 = 0.9, t2 = 10, h2 = 0.8, tmem = 100
AMAT = 2 + 0.1 × (10 + 0.2 × 100) = 2 + 0.1 × 30 = 5 ns.
Now play with it
- →Click the three tabs to see which levels are visited and the running time.
- →Change the hit ratios in the calculator and watch AMAT respond.
- CPU sends the address to L1
- L1 lookup costs t1: HIT, data returned. AMAT contribution = t1
Hierarchical: t1 + m1(t2 + m2·tm). Simultaneous: h1t1 + (1−h1)h2t2 + (1−h1)(1−h2)tm. Use the one the question wording implies.
Check yourself
t1 = 2 ns, h1 = 0.9, t2 = 10 ns, h2 = 0.8, memory 100 ns (hierarchical). What is AMAT?
Exam notes (the dense version)
L1 small & fast, L2 bigger & slower. Local miss rate of L2 is relative to L2 accesses; global = product.
Previous year questions: Cache Memory (22)
PYQs are grouped by topic, so this is the full set for Cache Memory; try the ones that match this subtopic. Open all 22 on the topic page.