Measuring CPU core-to-core latency

We measure the latency it takes for a CPU to send a message to another CPU via its cache coherence protocol.

By pinning two threads on two different CPU cores, we can get them to do a bunch of compare-exchange operation, and measure the latency.

How to run:

$ cargo install core-to-core-latency
$ core-to-core-latency

Single socket results

Intel Xeon Phi 7210, 64 Cores, Knights Landing, 2016-Q2

Data provided by Concyclics.

HiSilicon Kunpeng 920-6426, 64 cores, ARMv8.2-A, 2019-Q1

Data provided by Concyclics.

Intel Xeon Gold 6242, 16 Cores, Cascade Lake, 2019-Q2

Data provided by Concyclics.

AMD Ryzen 9 7950X, 16 Cores, zen4, 2022-Q3

Data provided by zamadatix.

Data provided by zamadatix.

AMD EPYC 7773X, 64 Cores, Milan-X, 2022-Q1

Data provided by SchrodingerZhu.

Loongson 3A5000HV, 2.5GHz, 4 Cores, 2021-Q3

Data provided by Glavo.

Intel Core i9-12900K, 8P+8E Cores, Alder Lake, 12th gen, 2021-Q4

Data provided by bizude.

This CPU has 8 performance cores, and 2 groups of 4 efficient cores. We see CPU=8 with fast access to all other cores.

Intel Core i9-9900K, 3.60GHz, 8 Cores, Coffee Lake, 8th gen, 2018-Q4

Data provided by nviennot.

Intel Core i7-1165G7, 2.80GHz, 4 Cores, Tiger Lake, 11th gen, 2020-Q3

Data provided by Jonas Wunderlich.

Intel Core i7-6700K, 4.00GHz, 4 Cores, Skylake, 6th gen, 2015-Q3

Data provided by CanIGetaPR.

Intel Core i5-10310U, 4 Cores, Comet Lake, 10th gen, 2020-Q2

Data provided by Ashley Sommer.

Intel Core i5-4590, 3.30GHz, 4 Cores, Haswell, 4th gen, 2014-Q2

Data provided by Felipe Lube de Bragança.

Apple M1 Pro, 6P+2E Cores, 2021-Q4

Data provided by Aditya Sharma.

We see the two efficent cores clustered together with a latency of 53ns, then two groups of 3 performance cores, with a latency of 40ns. Cross-group communication is slow at ~145ns, which is a latency typically seen in multi-socket configurations.

Intel Xeon Platinum 8375C, 2.90GHz 32 Cores, Ice Lake, 3rd gen, 2021-Q2

From an AWS c6i.metal machine.

Intel Xeon Platinum 8275CL, 3.00GHz 24 Cores, Cascade Lake, 2nd gen, 2019-Q2

From an AWS c5.metal machine.

Intel Xeon E5-2695 v4, 2.10GHz 18 Cores, Broadwell, 5th gen, 2016-Q1

From a machine provided by GTHost

AMD EPYC 7R13, 48 Cores, Milan, 3rd gen, 2021-Q1

From an AWS c6a.metal machine.

We can see cores arranged in 6 groups of 8 in which latency is excellent within (23ns). When data crosses groups, the latency jumps to around 110ns. Note, that the last 3 groups have a better cross-group latency than the first 3 (~90ns).

AMD Ryzen Threadripper 3960X, 3.80GHz, 24 Cores, Zen 2, 3rd Gen, 2019-Q4

Data provided by Mathias Siegel.

We see the CPUs in 8 groups of 3, and better performance for CPUS in the group [13,24].

AMD Ryzen Threadripper 1950X, 3.40GHz, 16 Cores, Zen, 1st Gen, 2017-Q3

Data provided by Jakub Okoński

We see the CPUs in 4 groups of 4, and better performance for CPUS in the group [9,16].

AMD Ryzen 9 5950X, 3.40GHz 16 Cores, Zen3, 4th gen, 2020-Q1

Data provided by John Schoenick.

We can see two groups of 8 cores with latencies of 17ns intra-group, and 85ns inter-group.

AMD Ryzen 9 5900X, 3.40GHz, 12 Cores, Zen3, 4th gen, 2020-Q4

Data provided by Scott Markwell.

We see two groups of 6 cores with latencies of 16ns intra-group and 84ns inter-group.

AMD Ryzen 7 5800U, 1.9GHz up to 4.4GHz, 8 Cores, Zen3, 4th gen, 2021-Q4

Data provided by George Melikov.

AMD Ryzen 7 5700X, 3.40GHz, 8 Cores, Zen3, 4th gen, 2022-Q2

Data provided by Ashley Sommer.

AMD Ryzen 7 2700X, 3.70GHz, 8 Cores, Zen+, 2nd gen, 2018-Q3

Data provided by David Hoppenbrouwers.

We can see 2 groups of 4 cores with latencies of 24ns intra-group, and 92ns inter-group.

AMD Ryzen 9 5900HX, 3.3GHz, 8 Cores, Zen3, 4th gen, 2021-Q1

Data provided by r4nd0m1z3r.

AWS Graviton3, 64 Cores, Arm Neoverse, 3rd gen, 2021-Q4

From an AWS c7g.16xlarge machine.

AWS Graviton2, 64 Cores, Arm Neoverse, 2nd gen, 2020-Q1

From an AWS c6gd.metal machine.

Sun/Oracle SPARC T4, 2.85GHz, 8 cores, 2011-Q3

Data provided by Kokoa van Houten.

IBM Power7, 3.3GHz, 8 Cores, 2010-Q1

Data provided by Kokoa van Houten.

Dual sockets results

The following shows dual-socket configuration latency where one CPU on the first socket sends a message to another CPU on the second socket. The number in parenthesis next to the latency denotes the slowdown compared to single socket.

Dual Intel Xeon Gold 6242, 16 Cores, Cascade Lake, 2019-Q2

Data provided by Concyclics.

Dual Intel Xeon Platinum 8375C, 2.90GHz 32 Cores, Ice Lake, 3rd gen, 2021-Q2

From an AWS c6i.metal machine.

Dual Intel Xeon Platinum 8275CL, 3.00GHz 24 Cores, Cascade Lake, 2nd gen, 2019-Q2

From an AWS c5.metal machine.

Dual Intel Xeon E5-2695 v4, 2.10GHz 18 Cores, Broadwell, 5th gen, 2016-Q1

From a machine provided by GTHost

Dual AMD EPYC 7R13, 48 Cores, Milan, 3rd gen, 2021-Q1

From an AWS c6a.metal machine.

This one is a bit odd. The single socket test for Socket 1 shows median latencies of 107ns cross-groups, but Socket 2 shows 200ns. It's 2x slower, very odd. The other platforms don't behave this way. In fact, the socket-to-socket latencies are than the core-to-core within Socket 2.

Anandtech have measured similar results on a Dual-Socket AMD EPYC 7763 and 7742.

Socket 2 does not behave similarly than Socket 1, it's twice as slow.

Sun/Oracle SPARC T4, 2.85GHz, 8 cores, 2011-Q3

Data provided by Kokoa van Houten.

Dual IBM Power7, 3.3GHz, 8 Cores, 2010-Q1

Data provided by Kokoa van Houten.

Hyper-threads

We measure the latency between two hyper-threads of the same core

The notebook results/results.ipynb contains the code to generate these graphs

How to use

First install Rust and gcc on linux, then:

$ cargo install core-to-core-latency
$ core-to-core-latency
Num cores: 10
Using RDTSC to measure time: false
Num round trips per samples: 1000
Num samples: 300
Showing latency=round-trip-time/2 in nanoseconds:

       0       1       2       3       4       5       6       7       8       9
  0
  1   52±6
  2   38±6    39±4
  3   39±5    39±6    38±6
  4   34±6    38±4    37±6    36±5
  5   38±5    38±6    38±6    38±6    37±6
  6   38±5    37±6    39±6    36±4    49±6    38±6
  7   36±6    39±5    39±6    37±6    35±6    36±6    38±6
  8   37±5    38±6    35±5    39±5    38±6    38±5    37±6    37±6
  9   48±6    39±6    36±6    39±6    38±6    36±6    41±6    38±6    39±6

Min  latency: 34.5ns ±6.1 cores: (4,0)
Max  latency: 52.1ns ±9.4 cores: (1,0)
Mean latency: 38.4ns

Contribute

Use core-to-core-latency 5000 --csv > output.csv to instruct the program to use 5000 iterations per sample to reduce the noise, and save the results.

It can be used in the jupter notebook results/results.ipynb for rendering graphs.

Create a GitHub issue with the generated output.csv file and I'll add your results.

License

This software is licensed under the MIT license