Jetson TX2 (Tegra 186) - ssvb/tinymembench GitHub Wiki
$ cat /proc/cpuinfo processor : 0 model name : ARMv8 Processor rev 3 (v8l) BogoMIPS : 62.50 Features : fp asimd evtstrm aes pmull sha1 sha2 crc32 CPU implementer : 0x41 CPU architecture: 8 CPU variant : 0x1 CPU part : 0xd07 CPU revision : 3 processor : 1 model name : ARMv8 Processor rev 0 (v8l) BogoMIPS : 62.50 Features : fp asimd evtstrm aes pmull sha1 sha2 crc32 CPU implementer : 0x4e CPU architecture: 8 CPU variant : 0x0 CPU part : 0x003 CPU revision : 0 MTS version : 37620520 processor : 2 model name : ARMv8 Processor rev 0 (v8l) BogoMIPS : 62.50 Features : fp asimd evtstrm aes pmull sha1 sha2 crc32 CPU implementer : 0x4e CPU architecture: 8 CPU variant : 0x0 CPU part : 0x003 CPU revision : 0 MTS version : 37620520 processor : 3 model name : ARMv8 Processor rev 3 (v8l) BogoMIPS : 62.50 Features : fp asimd evtstrm aes pmull sha1 sha2 crc32 CPU implementer : 0x41 CPU architecture: 8 CPU variant : 0x1 CPU part : 0xd07 CPU revision : 3 processor : 4 model name : ARMv8 Processor rev 3 (v8l) BogoMIPS : 62.50 Features : fp asimd evtstrm aes pmull sha1 sha2 crc32 CPU implementer : 0x41 CPU architecture: 8 CPU variant : 0x1 CPU part : 0xd07 CPU revision : 3 processor : 5 model name : ARMv8 Processor rev 3 (v8l) BogoMIPS : 62.50 Features : fp asimd evtstrm aes pmull sha1 sha2 crc32 CPU implementer : 0x41 CPU architecture: 8 CPU variant : 0x1 CPU part : 0xd07 CPU revision : 3 $ cat /sys/devices/system/cpu/cpu*/cpufreq/scaling_cur_freq 1881600 2035200 2035200 1881600 1881600 1881600
Denver:
$ taskset 6 ./tinymembench
tinymembench v0.4.9 (simple benchmark for memory throughput and latency)
==========================================================================
== Memory bandwidth tests ==
== ==
== Note 1: 1MB = 1000000 bytes ==
== Note 2: Results for 'copy' tests show how many bytes can be ==
== copied per second (adding together read and writen ==
== bytes would have provided twice higher numbers) ==
== Note 3: 2-pass copy means that we are using a small temporary buffer ==
== to first fetch data into it, and only then write it to the ==
== destination (source -> L1 cache, L1 cache -> destination) ==
== Note 4: If sample standard deviation exceeds 0.1%, it is shown in ==
== brackets ==
==========================================================================
C copy backwards : 7518.8 MB/s
C copy backwards (32 byte blocks) : 7517.0 MB/s (0.1%)
C copy backwards (64 byte blocks) : 7428.9 MB/s
C copy : 7597.6 MB/s (0.1%)
C copy prefetched (32 bytes step) : 7457.8 MB/s (0.1%)
C copy prefetched (64 bytes step) : 7720.9 MB/s (0.5%)
C 2-pass copy : 4516.6 MB/s
C 2-pass copy prefetched (32 bytes step) : 4234.2 MB/s (0.1%)
C 2-pass copy prefetched (64 bytes step) : 4545.2 MB/s
C fill : 14308.2 MB/s
C fill (shuffle within 16 byte blocks) : 14304.2 MB/s
C fill (shuffle within 32 byte blocks) : 14258.6 MB/s
C fill (shuffle within 64 byte blocks) : 14263.8 MB/s
---
standard memcpy : 6688.2 MB/s
standard memset : 14271.0 MB/s (0.2%)
---
NEON LDP/STP copy : 7103.8 MB/s
NEON LDP/STP copy pldl2strm (32 bytes step) : 7097.7 MB/s
NEON LDP/STP copy pldl2strm (64 bytes step) : 7101.0 MB/s (0.3%)
NEON LDP/STP copy pldl1keep (32 bytes step) : 7101.3 MB/s (0.2%)
NEON LDP/STP copy pldl1keep (64 bytes step) : 7099.5 MB/s (1.1%)
NEON LD1/ST1 copy : 7105.7 MB/s (0.2%)
NEON STP fill : 14776.3 MB/s
NEON STNP fill : 22990.8 MB/s (0.3%)
ARM LDP/STP copy : 6918.8 MB/s (0.3%)
ARM STP fill : 14279.9 MB/s (0.3%)
ARM STNP fill : 17720.9 MB/s
==========================================================================
== Framebuffer read tests. ==
== ==
== Many ARM devices use a part of the system memory as the framebuffer, ==
== typically mapped as uncached but with write-combining enabled. ==
== Writes to such framebuffers are quite fast, but reads are much ==
== slower and very sensitive to the alignment and the selection of ==
== CPU instructions which are used for accessing memory. ==
== ==
== Many x86 systems allocate the framebuffer in the GPU memory, ==
== accessible for the CPU via a relatively slow PCI-E bus. Moreover, ==
== PCI-E is asymmetric and handles reads a lot worse than writes. ==
== ==
== If uncached framebuffer reads are reasonably fast (at least 100 MB/s ==
== or preferably >300 MB/s), then using the shadow framebuffer layer ==
== is not necessary in Xorg DDX drivers, resulting in a nice overall ==
== performance improvement. For example, the xf86-video-fbturbo DDX ==
== uses this trick. ==
==========================================================================
NEON LDP/STP copy (from framebuffer) : 1579.4 MB/s (3.3%)
NEON LDP/STP 2-pass copy (from framebuffer) : 1991.1 MB/s (8.7%)
NEON LD1/ST1 copy (from framebuffer) : 1812.0 MB/s (1.7%)
NEON LD1/ST1 2-pass copy (from framebuffer) : 1915.3 MB/s (1.2%)
ARM LDP/STP copy (from framebuffer) : 396.7 MB/s (9.6%)
ARM LDP/STP 2-pass copy (from framebuffer) : 280.4 MB/s
==========================================================================
== Memory latency test ==
== ==
== Average time is measured for random memory accesses in the buffers ==
== of different sizes. The larger is the buffer, the more significant ==
== are relative contributions of TLB, L1/L2 cache misses and SDRAM ==
== accesses. For extremely large buffer sizes we are expecting to see ==
== page table walk with several requests to SDRAM for almost every ==
== memory access (though 64MiB is not nearly large enough to experience ==
== this effect to its fullest). ==
== ==
== Note 1: All the numbers are representing extra time, which needs to ==
== be added to L1 cache latency. The cycle timings for L1 cache ==
== latency can be usually found in the processor documentation. ==
== Note 2: Dual random read means that we are simultaneously performing ==
== two independent memory accesses at a time. In the case if ==
== the memory subsystem can't handle multiple outstanding ==
== requests, dual random read has the same timings as two ==
== single reads performed one after another. ==
==========================================================================
block size : single random read / dual random read
1024 : 0.0 ns / 0.0 ns
2048 : 0.0 ns / 0.0 ns
4096 : 0.0 ns / 0.0 ns
8192 : 0.0 ns / 0.0 ns
16384 : 0.0 ns / 0.0 ns
32768 : 0.0 ns / 0.0 ns
65536 : 0.0 ns / 0.0 ns
131072 : 4.0 ns / 7.1 ns
262144 : 6.0 ns / 10.1 ns
524288 : 7.0 ns / 11.5 ns
1048576 : 7.8 ns / 12.3 ns
2097152 : 37.5 ns / 63.6 ns
4194304 : 100.8 ns / 144.8 ns
8388608 : 137.1 ns / 172.3 ns
16777216 : 162.9 ns / 189.6 ns
33554432 : 176.3 ns / 198.1 ns
67108864 : 183.0 ns / 201.0 ns
Cortex-A57:
$ taskset 0x39 ./tinymembench
tinymembench v0.4.9 (simple benchmark for memory throughput and latency)
==========================================================================
== Memory bandwidth tests ==
== ==
== Note 1: 1MB = 1000000 bytes ==
== Note 2: Results for 'copy' tests show how many bytes can be ==
== copied per second (adding together read and writen ==
== bytes would have provided twice higher numbers) ==
== Note 3: 2-pass copy means that we are using a small temporary buffer ==
== to first fetch data into it, and only then write it to the ==
== destination (source -> L1 cache, L1 cache -> destination) ==
== Note 4: If sample standard deviation exceeds 0.1%, it is shown in ==
== brackets ==
==========================================================================
C copy backwards : 3576.3 MB/s (0.3%)
C copy backwards (32 byte blocks) : 3585.1 MB/s (0.2%)
C copy backwards (64 byte blocks) : 3576.9 MB/s (3.3%)
C copy : 3801.5 MB/s (0.4%)
C copy prefetched (32 bytes step) : 3975.2 MB/s (0.4%)
C copy prefetched (64 bytes step) : 3996.7 MB/s (0.3%)
C 2-pass copy : 3144.5 MB/s
C 2-pass copy prefetched (32 bytes step) : 3370.4 MB/s (0.2%)
C 2-pass copy prefetched (64 bytes step) : 3422.5 MB/s (0.3%)
C fill : 12865.2 MB/s
C fill (shuffle within 16 byte blocks) : 12871.6 MB/s
C fill (shuffle within 32 byte blocks) : 12870.2 MB/s
C fill (shuffle within 64 byte blocks) : 12865.8 MB/s
---
standard memcpy : 3697.4 MB/s (0.2%)
standard memset : 12868.1 MB/s
---
NEON LDP/STP copy : 3778.4 MB/s (0.3%)
NEON LDP/STP copy pldl2strm (32 bytes step) : 1661.3 MB/s
NEON LDP/STP copy pldl2strm (64 bytes step) : 1663.4 MB/s
NEON LDP/STP copy pldl1keep (32 bytes step) : 4014.8 MB/s (0.3%)
NEON LDP/STP copy pldl1keep (64 bytes step) : 4021.3 MB/s (0.2%)
NEON LD1/ST1 copy : 3775.6 MB/s (0.2%)
NEON STP fill : 12870.5 MB/s
NEON STNP fill : 12854.8 MB/s (0.4%)
ARM LDP/STP copy : 3764.7 MB/s (0.2%)
ARM STP fill : 12864.0 MB/s
ARM STNP fill : 12828.4 MB/s (0.7%)
==========================================================================
== Framebuffer read tests. ==
== ==
== Many ARM devices use a part of the system memory as the framebuffer, ==
== typically mapped as uncached but with write-combining enabled. ==
== Writes to such framebuffers are quite fast, but reads are much ==
== slower and very sensitive to the alignment and the selection of ==
== CPU instructions which are used for accessing memory. ==
== ==
== Many x86 systems allocate the framebuffer in the GPU memory, ==
== accessible for the CPU via a relatively slow PCI-E bus. Moreover, ==
== PCI-E is asymmetric and handles reads a lot worse than writes. ==
== ==
== If uncached framebuffer reads are reasonably fast (at least 100 MB/s ==
== or preferably >300 MB/s), then using the shadow framebuffer layer ==
== is not necessary in Xorg DDX drivers, resulting in a nice overall ==
== performance improvement. For example, the xf86-video-fbturbo DDX ==
== uses this trick. ==
==========================================================================
NEON LDP/STP copy (from framebuffer) : 763.9 MB/s
NEON LDP/STP 2-pass copy (from framebuffer) : 714.1 MB/s
NEON LD1/ST1 copy (from framebuffer) : 776.9 MB/s
NEON LD1/ST1 2-pass copy (from framebuffer) : 709.6 MB/s
ARM LDP/STP copy (from framebuffer) : 767.7 MB/s
ARM LDP/STP 2-pass copy (from framebuffer) : 719.1 MB/s
==========================================================================
== Memory latency test ==
== ==
== Average time is measured for random memory accesses in the buffers ==
== of different sizes. The larger is the buffer, the more significant ==
== are relative contributions of TLB, L1/L2 cache misses and SDRAM ==
== accesses. For extremely large buffer sizes we are expecting to see ==
== page table walk with several requests to SDRAM for almost every ==
== memory access (though 64MiB is not nearly large enough to experience ==
== this effect to its fullest). ==
== ==
== Note 1: All the numbers are representing extra time, which needs to ==
== be added to L1 cache latency. The cycle timings for L1 cache ==
== latency can be usually found in the processor documentation. ==
== Note 2: Dual random read means that we are simultaneously performing ==
== two independent memory accesses at a time. In the case if ==
== the memory subsystem can't handle multiple outstanding ==
== requests, dual random read has the same timings as two ==
== single reads performed one after another. ==
==========================================================================
block size : single random read / dual random read
1024 : 0.0 ns / 0.0 ns
2048 : 0.0 ns / 0.0 ns
4096 : 0.0 ns / 0.0 ns
8192 : 0.0 ns / 0.0 ns
16384 : 0.0 ns / 0.0 ns
32768 : 2.4 ns / 4.2 ns
65536 : 4.8 ns / 7.5 ns
131072 : 7.2 ns / 9.8 ns
262144 : 10.3 ns / 13.0 ns
524288 : 11.9 ns / 14.9 ns
1048576 : 13.1 ns / 16.1 ns
2097152 : 36.0 ns / 54.5 ns
4194304 : 93.4 ns / 133.9 ns
8388608 : 137.0 ns / 178.1 ns
16777216 : 159.1 ns / 201.8 ns
33554432 : 171.4 ns / 216.3 ns
67108864 : 187.3 ns / 242.7 ns