System Designer

Back to Reference

Latency Numbers

CPU, memory, storage, datacenter, and WAN baselines

Not Started

Loading...

Understanding latency hierarchies is crucial for system design. Each operation is orders of magnitude slower than the previous: CPU cache access takes nanoseconds, memory takes ~100ns, SSD takes microseconds, network takes milliseconds, and disk seeks take ~10ms. These numbers help you identify bottlenecks and optimize at the right level.

CPU / Memory

L1 Cache Reference

CPU‑local; near zero

0.5 ns

Branch Mispredict

Pipeline flush

5 ns

L2 Cache Reference

Shared per core cluster

7 ns

Mutex Lock/Unlock

Uncontended fast‑path

25 ns

Main Memory Reference

NUMA/DDR; locality matters

100 ns

Storage

Compress 1 KB (zippy)

CPU bound; varies with level

3 μs

Read 4 KB from SSD

Random I/O

150 μs

Read 1 MB from Memory

Sequential memcpy

250 μs

Read 1 MB from SSD

NVMe, QD impacts

1 ms

HDD Disk Seek

Rotational latency

10 ms

Read 1 MB from HDD

Sustained read

20 ms

Network

Send 1 KB over 1 Gbps

Serialization only

10 μs

Round trip in datacenter

Same DC, multiple hops

500 μs

Packet California → Netherlands

Speed‑of‑light floor

150 ms

Intra‑DC RTT

0.5ms

Cross‑Region RTT

80ms

Intercontinental RTT

150ms

📝 Test Your Knowledge

5 questions • Progress: 0/5