cpus.me

Latency numbers

Jeff Dean’s list with current measurements, more rows, and the source of each figure.

1 ns1 us1 ms1 sL1 hitMain memoryNVMe random readHDD random readSame-zone round tripIntercontinental round tripFrontier LLM first token60 Hz frameAttention lost
  • nanoseconds
  • microseconds
  • milliseconds
  • seconds

Updated 17 October 2026 / 23 sources / scale is logarithmic, 0.1 ns to 100 s

What these numbers are

Typical figures on an unloaded system, rounded to be memorised. Use them to check whether a design is off by an order of magnitude, not to predict production latency.

What they leave out

Queueing, contention and tail latency. Storage and network figures also vary widely across vendors and instance types; where a number matters, measure your own.

Processor and memory

Nanoseconds depend on clock speed, so cycle counts are given where the source has them.

Operation
Latency
Notes
Src
One CPU cycle at 5 GHz
0.2 ns
L1 data cache hit
1 ns
4 cycles on Zen 5 and Lion Cove, 5 on Raptor Cove, all at 48 KiB.
6
L2 cache hit
2 to 4 ns
14 cycles on Zen 5 (1 MiB), 16 on Raptor Cove (2 MiB), 17 on Lion Cove (3 MiB).
6
Branch mispredict
3 to 5 ns
15 to 16 cycles on Zen 3. The 2012 list gives 5 ns.
5,1
Non-cryptographic hash, 64 B
10 ns
5 GiB/s.
3
L3 or last-level cache hit
8 to 25 ns
About 47 cycles on Zen 5, 68 on Raptor Cove, 83 on Lion Cove. Ice Lake server measures 22.5 ns.
6,7
Random memory read, many in flight
20 ns
3 GiB/s.
3
Uncontended mutex lock and unlock
25 ns
1
Main memory, pointer chase
70 to 130 ns
Just over 70 ns on a desktop Ryzen 9 9950X with DDR5-6000; over 128 ns on a Ryzen AI 9 HX 370 laptop.
8,9
Cryptographic hash, 64 B
100 ns
1 GiB/s.
3
System call
300 ns
3
GPU global memory read (HBM)
375 to 400 ns
A100 566 cycles, H800 656 cycles, at the boost clocks given in the paper.
10
Sequential memory read, 1 MiB, all threads
5 us
200 GiB/s aggregate.
3
Context switch
1 to 10 us
Bendersky measured 1.2 to 2.2 us of direct cost; napkin-math budgets 10 us.
3,11
Sequential memory read, 1 MiB, one thread
50 us
20 GiB/s.
3

Storage

napkin-math rows were measured through a filesystem on a RAID0 of cloud local SSDs. Named drives are vendor specifications.

Operation
Latency
Notes
Src
NVMe sequential read, 8 KiB
1 us
8 GiB/s.
3
Sequential write, 8 KiB, no fsync
2 us
3 GiB/s, into the page cache.
3
NVMe random read, 4 KiB, queue depth 1
20 to 75 us
Micron 7500: 70 us typical, 80 us at p99. Seagate Nytro 5350H: 75 us.
12,13,14
NVMe random read, 8 KiB
100 us
70 MiB/s.
3
SATA SSD random read, 4 KiB
100 to 200 us
14
NVMe sequential write, 8 KiB, with fsync
300 us
30 MiB/s.
3
HDD sequential read, 1 MiB
4 ms
250 MiB/s.
3
HDD random read, 8 KiB
10 ms
0.7 MiB/s.
3
Blob storage GET returning 304
30 ms
S3 or GCS conditional request.
3
Blob storage GET, 128 KiB, one connection
80 ms
Large objects stream at 100 to 200 MiB/s on one connection, 2 to 5 GiB/s with concurrent range reads.
3
Blob storage LIST
100 ms
One 1000-key page.
3
Blob storage PUT, 128 KiB, one connection
200 ms
Priced about 12x a GET per request.
3

Network

Round trip times unless stated otherwise.

Operation
Latency
Notes
Src
Fibre propagation, 1 km, one way
4.9 us
Refractive index about 1.468, so light moves at roughly 204,000 km/s.
15,16
Put 1 KiB on the wire at 1 Gbit/s
8.2 us
8,192 bits at 10^9 bit/s, before framing. The 2012 list gives 10 us.
1
Proxy hop (Envoy, nginx, HAProxy)
50 us
3
TCP echo, 32 KiB
50 us
500 MiB/s, over loopback.
3
Round trip, same zone or VPC
250 us
25 GiB/s on premium networking.
3
Round trip, same region
250 us
2 GiB/s.
3
Round trip, same datacentre
500 us
The 2012 figure.
1
Redis, Memcached or MySQL query
500 us
3
US Central to US East
25 ms
3
US Central to US West
40 ms
3
US East to US West
60 ms
3
EU West to US East
80 ms
3
EU West to US Central
100 ms
3
California to Netherlands and back
150 ms
The 2012 figure.
1
EU West to Singapore
160 ms
3
US West to Singapore
180 ms
About 25 MiB/s between regions.
3

Software operations

Single core, data in memory. Per mebibyte, except signatures, which are per 59-byte message.

Operation
Latency
Notes
Src
Ed25519 sign
7.3 us
27,800 cycles on Zen 4 at 3.8 GHz.
17
Ed25519 verify
31 us
117,900 cycles on Zen 4 at 3.8 GHz.
17
AES-256-GCM encrypt, 1 MiB
155 us
0.56 cycles per byte on Zen 4 at 3.8 GHz.
19
Non-cryptographic hash, 1 MiB
200 us
5 GiB/s.
3
SHA-256, 1 MiB
560 us
2.02 cycles per byte on Zen 4 at 3.8 GHz.
18
Decompress 1 MiB
1 ms
1 GiB/s.
3
Serialise 1 MiB, flat wire format
1 ms
1 GiB/s. Simple wire formats that dump bytes.
3
Compress 1 MiB
2 ms
500 MiB/s. Each extra 1x of compression ratio costs about 10x in speed.
3
Sort 1 MiB of 64-bit integers
2 ms
500 MiB/s.
3
Serialise 1 MiB, general purpose
10 ms
100 MiB/s. JSON is in this class.
3

Language model inference

Added to the gist in May 2026 without a stated source. These will date fastest.

Operation
Latency
Notes
Src
Local model, generate one token
15 ms
Small model on a consumer GPU.
1
Frontier model, generate one token
20 ms
Hosted model. 50 tokens per second.
1
Local model, time to first token
75 ms
Small model, short prompt.
1
Local model on CPU, generate one token
100 ms
Small model, no GPU.
1
Specialised inference hardware, time to first token
250 ms
1
Frontier model, time to first token
1 s
Short prompt, no cache.
1
Frontier model, short response
3 s
About 100 output tokens.
1
Frontier model, prefill 100K tokens
10 s
No cache.
1
Frontier model, reasoning response
30 s
One call, with thinking.
1

Human perception

Nielsen's 1993 summary of limits that go back to 1968, plus the frame budget from RAIL.

Operation
Latency
Notes
Src
One frame at 60 Hz
16.7 ms
Budget about 10 ms of your own work; the browser needs the rest.
21,22
Feels instantaneous
100 ms
Handle the input within 50 ms to render something by 100 ms.
20,22
Keeps the flow of thought
1 s
Past this, attention drifts off the task.
20,21
Limit of attention
10 s
Past this, people turn to other tasks while they wait.
20

What follows from the table

1 ms of round trip buys about 50 km

Fibre costs 4.9 us per km each way, so a 1 ms round trip reaches about 100 km of fibre. Internet paths behave like about 2x the great-circle distance, so halve it.

Durability costs about 150x

The same 8 KiB write is 2 us buffered and 300 us with fsync.

Random access costs about 100x on flash

Sequential 8 KiB from NVMe is 1 us; random is 100 us. On a spinning disk the ratio is about 350x: 250 MiB/s sequential against 0.7 MiB/s random.

Utilisation multiplies latency

For an M/M/1 queue, response time is service time divided by (1 - utilisation): 2x at 50% load, 5x at 80%, 20x at 95%.

Sources

Entries marked further reading back no row.

  1. 1
    Latency Numbers Every Programmer Should Know

    Jeff Dean's figures, kept by Jonas Boner. Language model rows added in May 2026.

  2. 2
    Teach Yourself Programming in Ten Years, answers table further reading

    Peter Norvig, 2001. Its timing table was replaced around 2010 with Jeff Dean's figures.

  3. 3
    napkin-math

    Simon Eskildsen. The single-host rows were re-measured on 8 March 2026 on GCP c4-standard-48-lssd instances (Intel Xeon 6985P-C, 24 physical cores, 180 GB RAM, Ubuntu 22.04.5). Most throughput figures on this page come from here.

  4. 4
    Interactive latency page further reading

    Colin Scott. The same numbers projected across years from hardware trends; a model, not measurements.

  5. 5
    7-cpu.com: AMD Zen 3

    Branch misprediction penalty 15 to 16 cycles on a mOp cache hit.

  6. 6
    Intel's Lion Cove P-core and gaming workloads

    Chips and Cheese. Cache latency in cycles for Lion Cove, Raptor Cove and Zen 5.

  7. 7
    Measuring cache performance with Intel Memory Latency Checker (MLC)

    Zhiwei Chen. Ice Lake server (Xeon 8380): L1 1.5 ns, L2 4.1 ns, L3 22.5 ns.

  8. 8
    Ryzen 9950X, Zen 5 on desktop

    Chips and Cheese. Just over 70 ns of memory latency with DDR5-6000.

  9. 9
    Zen 5 variants, clock for clock

    Chips and Cheese. Pointer chase through a 1 GB array measured over 128 ns on a Ryzen AI 9 HX 370.

  10. 10
    Dissecting the NVIDIA Hopper architecture through microbenchmarking and multiple level analysis

    arXiv 2501.12084. Global memory latency in cycles for A100, H800 and RTX 4090.

  11. 11
    Measuring context switching and memory overheads for Linux threads

    Eli Bendersky. 1.2 to 1.5 us per switch pinned to one core, about 2.2 us unpinned.

  12. 12
    Micron 7500 SSD technical product specification

    4 KiB at queue depth 1, measured with fio: read 70 us typical and 80 us at p99, write 15 us typical.

  13. 13
    Seagate Nytro 5550 and 5350 data sheet

    Nytro 5350H average 4 KiB queue-depth-1 latency: read 75 us, write 12 us.

  14. 14
    NVMe latency, typical numbers

    simplyblock. NVMe 20 to 70 us, SATA SSD 100 to 200 us, HDD 5 to 10 ms.

  15. 15
    Calculating optical fiber latency

    m2optics. 4.9 us per km at a refractive index of 1.47.

  16. 16
    Dissecting latency in the Internet's fiber infrastructure

    arXiv 1811.10737. Uses 204,000 km/s; measured Internet latency is a median 2.1x the great-circle distance at that speed.

  17. 17
    eBATS: signature measurements on hertz

    Zen 4 Ryzen 7 7700 at 3.8 GHz, supercop-20261004. Ed25519 median 27,806 cycles to sign and 117,924 to verify a 59-byte message.

  18. 18
    eBASH: hash measurements on hertz

    Zen 4 Ryzen 7 7700 at 3.8 GHz, supercop-20261004. SHA-256 at 2.02 cycles per byte on long messages.

  19. 19
    eBAEAD: authenticated encryption measurements on hertz

    Zen 4 Ryzen 7 7700 at 3.8 GHz, supercop-20261004. AES-256-GCM encrypt at 0.56 cycles per byte on long messages.

  20. 20
    Response Times: The 3 Important Limits

    Jakob Nielsen, 1993. 0.1 s, 1 s and 10 s, after Miller (1968) and Card et al. (1991).

  21. 21
    Introducing RAIL: a user-centric model for performance

    Paul Irish and Paul Lewis, 2015. Adds the 16 ms animation frame to Nielsen's thresholds.

  22. 22
    Measure performance with the RAIL model

    web.dev. 10 ms of work per frame because the browser needs about 6 ms; process input within 50 ms to respond within 100 ms.

  23. 23
    cloudping.co further reading

    Measured round trip times between AWS regions.