[Parallel Computer archi]

Lecture3: Multi-core CPU

Why parallelism?

• ILP(Instruction Level Parallelism) doesn’t provide CPU performance improvement (ILP ~= 4)
• Parallelism is “primary” way to increase application performance
  • data parallelism, thread parallelism is main stream

Two key concepts that driven modern computer

1. Parallel execution
2. Efficiently accessing memory

1. Parallelism

Idea #1: Increasing # of transistor to add more cores

Rathre than sophisticated logic that help single-thread performance

Just duplicate core!!

• each core is slower than when running a single instruction stream
• But, Due to multi-core, Total speed-up increased

If # of core increased, execution context(register file) for each core should be smaller.

Beacause we have same die area! → Individual register file should be smaller

Idea #2: Add ALUs to increase computer power (SIMD paradigm)

Amortize cost/complexity of manging instruction stream

Single Instruction, Multiple Data
- Same instruction broadcast to all ALUs
- And executed in parallel on all ALUs
- ex) vector program

Idea #3: Idea #1 + Idea #2 ⇒ Mult “SIMD” core processor