The Ram Read / Wirte Speed ?

[xalius]

Hi Joe, can you fix your thread title please? While I tested RAM using ssvb's tinymembench on Rock64 (4GB LPDDR3-1600), there are also some iozone results of an A1 Sandisk card in the SD-Card thread (https://forum.pine64.org/showthread.php?...8#pid28378).

Code:

========================================================================== == 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                                     :   1433.8 MB/s (0.7%) C copy backwards (32 byte blocks)                    :   1462.1 MB/s C copy backwards (64 byte blocks)                    :   1359.9 MB/s (0.4%) C copy                                               :   1371.3 MB/s (0.4%) C copy prefetched (32 bytes step)                    :   1295.9 MB/s C copy prefetched (64 bytes step)                    :   1461.6 MB/s (0.2%) C 2-pass copy                                        :   1660.7 MB/s (0.1%) C 2-pass copy prefetched (32 bytes step)             :   1223.4 MB/s C 2-pass copy prefetched (64 bytes step)             :   1178.7 MB/s (0.1%) C fill                                               :   5694.9 MB/s C fill (shuffle within 16 byte blocks)               :   5694.3 MB/s C fill (shuffle within 32 byte blocks)               :   5694.4 MB/s C fill (shuffle within 64 byte blocks)               :   5696.0 MB/s --- standard memcpy                                      :   1330.3 MB/s standard memset                                      :   5696.0 MB/s --- NEON LDP/STP copy                                    :   1532.2 MB/s NEON LDP/STP copy pldl2strm (32 bytes step)          :   1272.3 MB/s (0.5%) NEON LDP/STP copy pldl2strm (64 bytes step)          :   1476.1 MB/s NEON LDP/STP copy pldl1keep (32 bytes step)          :   1677.4 MB/s NEON LDP/STP copy pldl1keep (64 bytes step)          :   1679.9 MB/s NEON LD1/ST1 copy                                    :   1515.7 MB/s (0.2%) NEON STP fill                                        :   5696.7 MB/s NEON STNP fill                                       :   2277.0 MB/s ARM LDP/STP copy                                     :   1531.5 MB/s ARM STP fill                                         :   5696.4 MB/s ARM STNP fill                                        :   2336.4 MB/s (1.1%) ========================================================================== == 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)                 :    306.2 MB/s NEON LDP/STP 2-pass copy (from framebuffer)          :    288.2 MB/s NEON LD1/ST1 copy (from framebuffer)                 :     80.5 MB/s NEON LD1/ST1 2-pass copy (from framebuffer)          :     79.4 MB/s ARM LDP/STP copy (from framebuffer)                  :    157.8 MB/s ARM LDP/STP 2-pass copy (from framebuffer)           :    152.9 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 :    0.1 ns          /     0.1 ns     65536 :    5.3 ns          /     9.0 ns    131072 :    8.1 ns          /    12.7 ns    262144 :   11.3 ns          /    16.4 ns    524288 :   66.5 ns          /   104.9 ns   1048576 :  101.0 ns          /   142.5 ns   2097152 :  119.3 ns          /   157.9 ns   4194304 :  133.1 ns          /   169.8 ns   8388608 :  141.1 ns          /   177.0 ns  16777216 :  146.9 ns          /   182.6 ns  33554432 :  150.9 ns          /   186.5 ns  67108864 :  163.6 ns          /   207.0 ns