<aside> 🚅

Empirical evaluation metrics: performance (this course's focus), Cost, Power, Reliability

• Combined metrics often matter more than individual ones: Performance/cost (MIPS/$), Performance/power (MIPS/W) — the latter is key for accelerators
• These metrics drive design decisions and purchasing decisions </aside>

Latency vs Throughput

• Latency (execution time): time required to finish one fixed task
• Throughput (bandwidth): number of tasks completed per unit time
  • Throughput can exploit parallelism; latency cannot. The two are often contradictory, so pick the definition that matches your goal (usually throughput).

Example

• Car: capacity 5, speed 60 mph → Latency = 10 min; Throughput = 30 people/hr (5 people × 6 trips/hr)
• Bus: capacity 60, speed 20 mph → Latency = 30 min; Throughput = 60 people/hr

Computing Performance ("X times faster" / "X% faster")

Latency(P,A) = Latency(P,B)/X
Throughput(P,A) = Throughput(P,B) × X

For X% faster:

Latency(P,A) = Latency(P,B)/(1 + X/100)
Throughput(P,A) = Throughput(P,B) × (1 + X/100)

What is “P” (the workload)?

• Actual target workload — most accurate but not portable, hard to pinpoint bottlenecks
• Representative benchmark programs — portable, repeatable, but may not match your exact use case
• Micro-benchmarks/kernels — easy to run, portable, but not representative of complex real-program behavior

Adding and averaging perforamnce numbers

You can add latencies, but not throughputs directly:

Latency(P1+P2, A) = Latency(P1,A) + Latency(P2,A)
Throughput(P1+P2,A) ≠ Throughput(P1,A) + Throughput(P2,A)

Combined throughput must be comptued from combined time: