M1Pro/M5Pro 芯片 MacBook Pro 性能评测

EpLiar 3月 13, 2026

前言

从 M1 Pro 购入到 M5 Pro 购入大概过去了四年五个月的时间，这期间迭代了很多款 M 系芯片。从 coding agent 出现后对于我的工作负载来说也还算够用，但最后还是决定购入 M5 Pro。趁着 M1 Pro 还没转手的这几天，撰写此文，以充分发挥 M1 Pro 的最后一点余温。

测试环境变量

一些介绍游戏打法的视频通常都会为了介绍容错性而故意/无意地进行防水，本文也不例外，测试环境在同一条件的情况下为 M5 Pro 的 MacBook Pro 新增了以下 Debuff：

落地扇会先吹到 M1 Pro 再吹它，也就是它在下风向
它比 M1 Pro 的后台多跑了包括但不限于微信、Telegram 和 Discord
它的电源是来自一根台式电脑 Type-A 接出来的 Type-C 数据线，M1 Pro 是 Anker 桌面充电器 (140W Port)
M1 Pro 关闭背光键盘，它打开

接下来，就看看测试结果吧

开始测试

Cinebench

测试使用版本：Cinebench 2026，gsha1sum 哈希值：b64b07dcd1fb0209c806350eed1dc70a4f29ea87 Cinebench2026_macOS.dmg

m1pro cinebench m5pro cinebench

型号	测试项目	分数	变化
M1Pro	GPU	9789
M5Pro	GPU	40175	+30386
M1Pro	CPU (Multiple Threads)	2634
M5Pro	CPU (Multiple Threads)	7702	+5068
M1Pro	CPU (Single Thread)	442
M5Pro	CPU (Single Thread)	738	+296
M1Pro	MP Ratio	5.96
M5Pro	MP Ratio	10.43	+4.47

型号

测试项目

分数

变化

M1Pro

GPU

9789

M5Pro

GPU

40175

+30386

M1Pro

CPU (Multiple Threads)

2634

M5Pro

CPU (Multiple Threads)

7702

+5068

M1Pro

CPU (Single Thread)

442

M5Pro

CPU (Single Thread)

738

+296

M1Pro

MP Ratio

5.96

M5Pro

MP Ratio

10.43

+4.47

LM Studio (大模型 Token/s 测试)

此测试将使用两个模型：google/gemma-3-4b 和 qwen/qwen3.5-9b，测试结果均遇到「停止原因：检测到 EOS Token」

测试提示词

Generate a numbered list starting from 1.

Each line must contain the number followed by a short random English sentence.

Example format:
1. The sky above the city slowly turned orange at sunset.
2. A small robot carefully sorted books on the wooden shelf.

Continue increasing the number and writing new sentences.

Do not stop. Continue until the maximum token limit is reached.

google/gemma-3-4b

型号	Token/s	Token 用量	TTFT
M1Pro	49.03	1260	0.83s
M5Pro	96.79	1379	0.24s

qwen/qwen3.5-9b

型号	Token/s	Token 用量	TTFT
M1Pro	16.2	2465	1.32s
M5Pro	34.78	3672	0.71s

ffmpeg 视频合成

HDR 10 样片来自于影视飓风的 YouTube：【HDR10】我们用一年拍下了这些样片…

ffmpeg 版本信息

ffmpeg version 8.0.1 Copyright (c) 2000-2025 the FFmpeg developers
  built with Apple clang version 17.0.0 (clang-1700.6.3.2)
  configuration: --prefix=/opt/homebrew/Cellar/ffmpeg/8.0.1_4 --enable-shared --enable-pthreads --enable-version3 --cc=clang --host-cflags= --host-ldflags= --enable-ffplay --enable-gpl --enable-libsvtav1 --enable-libopus --enable-libx264 --enable-libmp3lame --enable-libdav1d --enable-libvpx --enable-libx265 --enable-openssl --enable-videotoolbox --enable-audiotoolbox --enable-neon
  libavutil      60.  8.100 / 60.  8.100
  libavcodec     62. 11.100 / 62. 11.100
  libavformat    62.  3.100 / 62.  3.100
  libavdevice    62.  1.100 / 62.  1.100
  libavfilter    11.  4.100 / 11.  4.100
  libswscale      9.  1.100 /  9.  1.100
  libswresample   6.  1.100 /  6.  1.100
Universal media converter

下载/合成过程

$ yt-dlp -f 251 https://www.youtube.com/watch\?v\=itCWAiMWBgM --cookies-from-browser chrome
$ yt-dlp -f 701 https://www.youtube.com/watch\?v\=itCWAiMWBgM --cookies-from-browser chrome
$ ffmpeg -i "【HDR10】我们用一年拍下了这些样片… [itCWAiMWBgM].mp4" -i "【HDR10】我们用一年拍下了这些样片… [itCWAiMWBgM].webm" 我们用一年拍下了这些样片….mp4
Input #0, mov,mp4,m4a,3gp,3g2,mj2, from '【HDR10】我们用一年拍下了这些样片… [itCWAiMWBgM].mp4':
  Metadata:
    major_brand     : dash
    minor_version   : 0
    compatible_brands: iso6av01mp41
    creation_time   : 2026-02-16T01:59:50.000000Z
  Duration: 00:03:38.62, start: 0.000000, bitrate: 20907 kb/s
  Stream #0:0[0x1](und): Video: av1 (libdav1d) (Main) (av01 / 0x31307661), yuv420p10le(tv, bt2020nc/bt2020/smpte2084), 3840x2160, 20303 kb/s, 29.97 fps, 29.97 tbr, 30k tbn (default)
    Metadata:
      creation_time   : 2026-02-16T01:59:50.000000Z
      handler_name    : ISO Media file produced by Google Inc.
      vendor_id       : [0][0][0][0]
    Side data:
      Mastering Display Metadata, has_primaries:1 has_luminance:1 r(0.6800,0.3200) g(0.2650,0.6900) b(0.1500 0.0600) wp(0.3127, 0.3290) min_luminance=0.005000, max_luminance=1000.000000
Input #1, matroska,webm, from '【HDR10】我们用一年拍下了这些样片… [itCWAiMWBgM].webm':
  Metadata:
    encoder         : google/video-file
  Duration: 00:03:38.62, start: -0.007000, bitrate: 122 kb/s
  Stream #1:0(eng): Audio: opus, 48000 Hz, stereo, fltp, start -0.007000 (default)
Stream mapping:
  Stream #0:0 -> #0:0 (av1 (libdav1d) -> h264 (libx264))
  Stream #1:0 -> #0:1 (opus (native) -> aac (native))
Press [q] to stop, [?] for help
[libx264 @ 0xca6c31180] using SAR=1/1
[libx264 @ 0xca6c31180] using cpu capabilities: ARMv8 NEON DotProd I8MM
[libx264 @ 0xca6c31180] profile High 10, level 5.1, 4:2:0, 10-bit
[libx264 @ 0xca6c31180] x264 - core 165 r3222 b35605a - H.264/MPEG-4 AVC codec - Copyleft 2003-2025 - http://www.videolan.org/x264.html - options: cabac=1 ref=3 deblock=1:0:0 analyse=0x3:0x113 me=hex subme=7 psy=1 psy_rd=1.00:0.00 mixed_ref=1 me_range=16 chroma_me=1 trellis=1 8x8dct=1 cqm=0 deadzone=21,11 fast_pskip=1 chroma_qp_offset=-2 threads=22 lookahead_threads=3 sliced_threads=0 nr=0 decimate=1 interlaced=0 bluray_compat=0 constrained_intra=0 bframes=3 b_pyramid=2 b_adapt=1 b_bias=0 direct=1 weightb=1 open_gop=0 weightp=2 keyint=250 keyint_min=25 scenecut=40 intra_refresh=0 rc_lookahead=40 rc=crf mbtree=1 crf=23.0 qcomp=0.60 qpmin=0 qpmax=81 qpstep=4 mastering-display=G(13248,34500)B(7500,3000)R(34000,16000)WP(15635,16450)L(10000000,50) ip_ratio=1.40 aq=1:1.00
Output #0, mp4, to '我们用一年拍下了这些样片….mp4':
  Metadata:
    major_brand     : dash
    minor_version   : 0
    compatible_brands: iso6av01mp41
    encoder         : Lavf62.3.100
  Stream #0:0(und): Video: h264 (avc1 / 0x31637661), yuv420p10le(tv, bt2020nc/bt2020/smpte2084, progressive), 3840x2160 [SAR 1:1 DAR 16:9], q=2-31, 29.97 fps, 30k tbn (default)
    Metadata:
      encoder         : Lavc62.11.100 libx264
      creation_time   : 2026-02-16T01:59:50.000000Z
      handler_name    : ISO Media file produced by Google Inc.
      vendor_id       : [0][0][0][0]
    Side data:
      Mastering Display Metadata, has_primaries:1 has_luminance:1 r(0.6800,0.3200) g(0.2650,0.6900) b(0.1500 0.0600) wp(0.3127, 0.3290) min_luminance=0.005000, max_luminance=1000.000000
      cpb: bitrate max/min/avg: 0/0/0 buffer size: 0 vbv_delay: N/A
  Stream #0:1(eng): Audio: aac (LC) (mp4a / 0x6134706D), 48000 Hz, stereo, fltp, 128 kb/s (default)
    Metadata:
      encoder         : Lavc62.11.100 aac

测试结果：

M1 Pro：elapsed=0:06:33.60
M5 Pro：elapsed=0:02:45.39，提升 ≈ 57.98%

硬盘读写测试

左边为 M1 Pro，右边为 M5 Pro

m1pro disk test m5pro disk test

编译 Redis

本次使用的 Redis 源码为 https://github.com/redis/redis.git 仓库的 ee376cdc3cbf93428ee05829074fd83f33800f50 commit，make、gcc-15 版本如下：

$ make --version

GNU Make 3.81
Copyright (C) 2006  Free Software Foundation, Inc.
This is free software; see the source for copying conditions.
There is NO warranty; not even for MERCHANTABILITY or FITNESS FOR A
PARTICULAR PURPOSE.

This program built for i386-apple-darwin11.3.0

$ gcc-15 --version
gcc-15 (Homebrew GCC 15.2.0_1) 15.2.0
Copyright (C) 2025 Free Software Foundation, Inc.
This is free software; see the source for copying conditions.  There is NO
warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.

编译使用 make CC=gcc-15 -j$(sysctl -n hw.ncpu)，结果如下：

M1 Pro：31s
M5 Pro：14s，提升 ≈ 54.8%

M1 Pro 使用的编译命令是 make CC=gcc-15 -j8，因为它不认 hw.ncpu

内存持续带宽测试

该测试使用的仓库为：jeffhammond/STREAM ，commit hash: 6703f7504a38a8da96b353cadafa64d3c2d7a2d3，编译命令为 clang -O3 -march=native -DSTREAM_ARRAY_SIZE=80000000 stream.c -o stream，clang 版本为 Apple clang version 17.0.0 (clang-1700.6.4.2)

M1 Pro 测试结果

-------------------------------------------------------------
STREAM version $Revision: 5.10 $
-------------------------------------------------------------
This system uses 8 bytes per array element.
-------------------------------------------------------------
Array size = 80000000 (elements), Offset = 0 (elements)
Memory per array = 610.4 MiB (= 0.6 GiB).
Total memory required = 1831.1 MiB (= 1.8 GiB).
Each kernel will be executed 10 times.
 The *best* time for each kernel (excluding the first iteration)
 will be used to compute the reported bandwidth.
-------------------------------------------------------------
Your clock granularity/precision appears to be 1 microseconds.
Each test below will take on the order of 155800 microseconds.
   (= 155800 clock ticks)
Increase the size of the arrays if this shows that
you are not getting at least 20 clock ticks per test.
-------------------------------------------------------------
WARNING -- The above is only a rough guideline.
For best results, please be sure you know the
precision of your system timer.
-------------------------------------------------------------
Function    Best Rate MB/s  Avg time     Min time     Max time
Copy:           90799.6     0.014164     0.014097     0.014238
Scale:          85504.0     0.015104     0.014970     0.015421
Add:            77814.9     0.025042     0.024674     0.026577
Triad:          77767.6     0.025286     0.024689     0.028325
-------------------------------------------------------------
Solution Validates: avg error less than 1.000000e-13 on all three arrays
-------------------------------------------------------------

M5 Pro 测试结果

-------------------------------------------------------------
STREAM version $Revision: 5.10 $
-------------------------------------------------------------
This system uses 8 bytes per array element.
-------------------------------------------------------------
Array size = 80000000 (elements), Offset = 0 (elements)
Memory per array = 610.4 MiB (= 0.6 GiB).
Total memory required = 1831.1 MiB (= 1.8 GiB).
Each kernel will be executed 10 times.
 The *best* time for each kernel (excluding the first iteration)
 will be used to compute the reported bandwidth.
-------------------------------------------------------------
Your clock granularity/precision appears to be 1 microseconds.
Each test below will take on the order of 41908 microseconds.
   (= 41908 clock ticks)
Increase the size of the arrays if this shows that
you are not getting at least 20 clock ticks per test.
-------------------------------------------------------------
WARNING -- The above is only a rough guideline.
For best results, please be sure you know the
precision of your system timer.
-------------------------------------------------------------
Function    Best Rate MB/s  Avg time     Min time     Max time
Copy:          135221.7     0.009498     0.009466     0.009575
Scale:         126695.2     0.010163     0.010103     0.010318
Add:           131896.4     0.014612     0.014557     0.014814
Triad:         132414.7     0.014581     0.014500     0.014835
-------------------------------------------------------------
Solution Validates: avg error less than 1.000000e-13 on all three arrays
-------------------------------------------------------------

测试项	M1Pro	M5Pro	提升
Copy	90799.6 MB/s	135221.7 MB/s	48.92%
Scale	85504.0 MB/s	126695.2 MB/s	48.17%
Add	77814.9 MB/s	131896.4 MB/s	69.50%
Triad	77767.6 MB/s	132414.7 MB/s	70.27%

测试项

M1Pro

M5Pro

提升

Copy

90799.6 MB/s

135221.7 MB/s

48.92%

Scale

85504.0 MB/s

126695.2 MB/s

48.17%

Add

77814.9 MB/s

131896.4 MB/s

69.50%

Triad

77767.6 MB/s

132414.7 MB/s

70.27%

平均提升 ≈ 59.2%

rchase 内存随机访问测试

这里使用了 cyy 的项目：rchase，测试参数为 ./rchase 2048 10000000 ，连续测试五次

型号	最短时长	最长时长	平均值
M1Pro	142.65 ns	148.22 ns	143.922 ns
M5Pro	118.56 ns	120.61 ns	119.43 ns

型号

最短时长

最长时长

平均值

M1Pro

142.65 ns

148.22 ns

143.922 ns

M5Pro

118.56 ns

120.61 ns

119.43 ns

硬盘随机读写测试

该测试使用 fio-3.41 进行，测试命令为 fio --name=randrw --filename=testfile --size=8G --bs=4k --rw=randrw --rwmixread=50 --iodepth=64 --ioengine=posixaio --direct=1 --runtime=60 --time_based

M1 Pro 测试结果

randrw: (g=0): rw=randrw, bs=(R) 4096B-4096B, (W) 4096B-4096B, (T) 4096B-4096B, ioengine=posixaio, iodepth=64
fio-3.41
Starting 1 process
randrw: Laying out IO file (1 file / 8192MiB)
Jobs: 1 (f=1): [m(1)][100.0%][r=71.5MiB/s,w=71.0MiB/s][r=18.3k,w=18.2k IOPS][eta 00m:00s]
randrw: (groupid=0, jobs=1): err= 0: pid=98424: Sun Mar 15 16:36:19 2026
  read: IOPS=18.6k, BW=72.7MiB/s (76.3MB/s)(4365MiB/60001msec)
    slat (nsec): min=0, max=173000, avg=3237.49, stdev=5154.97
    clat (usec): min=4, max=9319, avg=373.36, stdev=541.11
     lat (usec): min=5, max=9333, avg=376.60, stdev=541.00
    clat percentiles (usec):
     |  1.00th=[   15],  5.00th=[   55], 10.00th=[  121], 20.00th=[  143],
     | 30.00th=[  157], 40.00th=[  172], 50.00th=[  184], 60.00th=[  200],
     | 70.00th=[  221], 80.00th=[  253], 90.00th=[ 1631], 95.00th=[ 1827],
     | 99.00th=[ 2008], 99.50th=[ 2073], 99.90th=[ 2245], 99.95th=[ 2638],
     | 99.99th=[ 6915]
   bw (  KiB/s): min=56928, max=81240, per=100.00%, avg=74562.91, stdev=3114.78, samples=119
   iops        : min=14232, max=20310, avg=18640.41, stdev=778.68, samples=119
  write: IOPS=18.6k, BW=72.7MiB/s (76.2MB/s)(4362MiB/60001msec); 0 zone resets
    slat (nsec): min=0, max=779000, avg=3317.50, stdev=5354.93
    clat (usec): min=5, max=9375, avg=467.41, stdev=560.51
     lat (usec): min=10, max=9377, avg=470.73, stdev=560.39
    clat percentiles (usec):
     |  1.00th=[   53],  5.00th=[  141], 10.00th=[  174], 20.00th=[  202],
     | 30.00th=[  223], 40.00th=[  243], 50.00th=[  265], 60.00th=[  289],
     | 70.00th=[  322], 80.00th=[  392], 90.00th=[ 1713], 95.00th=[ 1909],
     | 99.00th=[ 2114], 99.50th=[ 2180], 99.90th=[ 3490], 99.95th=[ 3654],
     | 99.99th=[ 7046]
   bw (  KiB/s): min=56968, max=80334, per=100.00%, avg=74526.04, stdev=3033.21, samples=119
   iops        : min=14242, max=20083, avg=18631.12, stdev=758.29, samples=119
  lat (usec)   : 10=0.15%, 20=0.93%, 50=1.72%, 100=2.25%, 250=56.58%
  lat (usec)   : 500=24.67%, 750=1.23%, 1000=0.16%
  lat (msec)   : 2=10.34%, 4=1.94%, 10=0.03%
  cpu          : usr=9.39%, sys=251.26%, ctx=7351546, majf=0, minf=9
  IO depths    : 1=0.1%, 2=0.1%, 4=0.2%, 8=62.3%, 16=37.4%, 32=0.0%, >=64=0.0%
     submit    : 0=0.0%, 4=100.0%, 8=0.0%, 16=0.0%, 32=0.0%, 64=0.0%, >=64=0.0%
     complete  : 0=0.0%, 4=96.6%, 8=3.2%, 16=0.2%, 32=0.0%, 64=0.0%, >=64=0.0%
     issued rwts: total=1117407,1116754,0,0 short=0,0,0,0 dropped=0,0,0,0
     latency   : target=0, window=0, percentile=100.00%, depth=64

Run status group 0 (all jobs):
   READ: bw=72.7MiB/s (76.3MB/s), 72.7MiB/s-72.7MiB/s (76.3MB/s-76.3MB/s), io=4365MiB (4577MB), run=60001-60001msec
  WRITE: bw=72.7MiB/s (76.2MB/s), 72.7MiB/s-72.7MiB/s (76.2MB/s-76.2MB/s), io=4362MiB (4574MB), run=60001-60001msec

M5 Pro 测试结果

randrw: (g=0): rw=randrw, bs=(R) 4096B-4096B, (W) 4096B-4096B, (T) 4096B-4096B, ioengine=posixaio, iodepth=64
fio-3.41
Starting 1 process
randrw: Laying out IO file (1 file / 8192MiB)
Jobs: 1 (f=1): [m(1)][100.0%][r=128MiB/s,w=128MiB/s][r=32.8k,w=32.8k IOPS][eta 00m:00s]
randrw: (groupid=0, jobs=1): err= 0: pid=791: Sun Mar 15 16:25:02 2026
  read: IOPS=31.8k, BW=124MiB/s (130MB/s)(7460MiB/60001msec)
    slat (nsec): min=0, max=106000, avg=1901.73, stdev=3024.24
    clat (usec): min=3, max=7214, avg=226.38, stdev=339.14
     lat (usec): min=3, max=7218, avg=228.28, stdev=339.09
    clat percentiles (usec):
     |  1.00th=[   11],  5.00th=[   83], 10.00th=[   93], 20.00th=[  103],
     | 30.00th=[  112], 40.00th=[  119], 50.00th=[  126], 60.00th=[  135],
     | 70.00th=[  145], 80.00th=[  163], 90.00th=[  235], 95.00th=[ 1270],
     | 99.00th=[ 1434], 99.50th=[ 1467], 99.90th=[ 1598], 99.95th=[ 2376],
     | 99.99th=[ 4359]
   bw (  KiB/s): min=103808, max=139601, per=100.00%, avg=127384.36, stdev=6568.27, samples=119
   iops        : min=25952, max=34900, avg=31845.85, stdev=1642.07, samples=119
  write: IOPS=31.8k, BW=124MiB/s (130MB/s)(7451MiB/60001msec); 0 zone resets
    slat (nsec): min=0, max=152000, avg=1973.04, stdev=3180.62
    clat (usec): min=4, max=7338, avg=266.26, stdev=353.25
     lat (usec): min=6, max=7338, avg=268.24, stdev=353.20
    clat percentiles (usec):
     |  1.00th=[   28],  5.00th=[  102], 10.00th=[  116], 20.00th=[  128],
     | 30.00th=[  137], 40.00th=[  147], 50.00th=[  157], 60.00th=[  169],
     | 70.00th=[  184], 80.00th=[  210], 90.00th=[  318], 95.00th=[ 1319],
     | 99.00th=[ 1483], 99.50th=[ 1532], 99.90th=[ 1729], 99.95th=[ 2540],
     | 99.99th=[ 4817]
   bw (  KiB/s): min=105354, max=139691, per=100.00%, avg=127254.03, stdev=6323.81, samples=119
   iops        : min=26338, max=34922, avg=31813.21, stdev=1580.94, samples=119
  lat (usec)   : 4=0.01%, 10=0.47%, 20=0.93%, 50=1.77%, 100=7.02%
  lat (usec)   : 250=78.39%, 500=2.66%, 750=0.04%, 1000=0.02%
  lat (msec)   : 2=8.61%, 4=0.07%, 10=0.01%
  cpu          : usr=12.65%, sys=249.01%, ctx=10271579, majf=0, minf=9
  IO depths    : 1=0.1%, 2=0.1%, 4=0.4%, 8=59.8%, 16=39.8%, 32=0.0%, >=64=0.0%
     submit    : 0=0.0%, 4=100.0%, 8=0.0%, 16=0.0%, 32=0.0%, 64=0.0%, >=64=0.0%
     complete  : 0=0.0%, 4=97.1%, 8=2.7%, 16=0.2%, 32=0.0%, 64=0.0%, >=64=0.0%
     issued rwts: total=1909637,1907506,0,0 short=0,0,0,0 dropped=0,0,0,0
     latency   : target=0, window=0, percentile=100.00%, depth=64

Run status group 0 (all jobs):
   READ: bw=124MiB/s (130MB/s), 124MiB/s-124MiB/s (130MB/s-130MB/s), io=7460MiB (7822MB), run=60001-60001msec
  WRITE: bw=124MiB/s (130MB/s), 124MiB/s-124MiB/s (130MB/s-130MB/s), io=7451MiB (7813MB), run=60001-60001msec

Web 渲染性能测试 (Speedometer3.0)

测试地址：Speedometer 3，着急出门测完才发现首页建议使用 Speedometer 3.1

M1 Pro 测试结果

TodoMVC-JavaScript-ES5  33.14 ± 1.87 (5.6%) ms
TodoMVC-JavaScript-ES6-Webpack-Complex-DOM  37.25 ± 4.77 (12.8%)  ms
TodoMVC-WebComponents 13.64 ± 1.38 (10.1%)  ms
TodoMVC-React-Complex-DOM 27.23 ± 2.08 (7.6%) ms
TodoMVC-React-Redux 28.60 ± 2.89 (10.1%)  ms
TodoMVC-Backbone  20.75 ± 1.11 (5.4%) ms
TodoMVC-Angular-Complex-DOM 25.02 ± 1.88 (7.5%) ms
TodoMVC-Vue 18.01 ± 2.42 (13.4%)  ms
TodoMVC-jQuery  95.79 ± 7.86 (8.2%) ms
TodoMVC-Preact-Complex-DOM  11.05 ± 1.03 (9.3%) ms
TodoMVC-Svelte-Complex-DOM  9.45  ± 0.90 (9.5%) ms
TodoMVC-Lit-Complex-DOM 12.53 ± 0.87 (6.9%) ms
NewsSite-Next 65.56 ± 3.96 (6.0%) ms
NewsSite-Nuxt 51.96 ± 3.42 (6.6%) ms
Editor-CodeMirror 20.09 ± 4.97 (24.7%)  ms
Editor-TipTap 55.42 ± 5.38 (9.7%) ms
Charts-observable-plot  38.99 ± 4.02 (10.3%)  ms
Charts-chartjs  36.75 ± 6.83 (18.6%)  ms
React-Stockcharts-SVG 63.34 ± 6.91 (10.9%)  ms

M5 Pro 测试结果

TodoMVC-JavaScript-ES5  19.20 ± 2.78 (14.5%)  ms
TodoMVC-JavaScript-ES6-Webpack-Complex-DOM  19.45 ± 1.93 (9.9%) ms
TodoMVC-WebComponents 8.76  ± 2.05 (23.4%)  ms
TodoMVC-React-Complex-DOM 17.41 ± 2.58 (14.8%)  ms
TodoMVC-React-Redux 17.57 ± 2.60 (14.8%)  ms
TodoMVC-Backbone  12.51 ± 2.32 (18.5%)  ms
TodoMVC-Angular-Complex-DOM 16.03 ± 1.85 (11.6%)  ms
TodoMVC-Vue 11.62 ± 1.70 (14.6%)  ms
TodoMVC-jQuery  61.52 ± 11.84 (19.2%) ms
TodoMVC-Preact-Complex-DOM  7.02  ± 0.80 (11.4%)  ms
TodoMVC-Svelte-Complex-DOM  6.38  ± 0.87 (13.6%)  ms
TodoMVC-Lit-Complex-DOM 8.17  ± 1.04 (12.7%)  ms
NewsSite-Next 37.84 ± 1.89 (5.0%) ms
NewsSite-Nuxt 31.13 ± 3.17 (10.2%)  ms
Editor-CodeMirror 12.93 ± 3.56 (27.5%)  ms
Editor-TipTap 36.06 ± 5.94 (16.5%)  ms
Charts-observable-plot  24.34 ± 3.79 (15.6%)  ms
Charts-chartjs  23.29 ± 5.89 (25.3%)  ms
React-Stockcharts-SVG 39.33 ± 4.11 (10.5%)  ms
Perf-Dashboard  16.76 ± 1.57 (9.4%) ms

测试项	M1Pro	M5Pro	提升
TodoMVC-JavaScript-ES5	33.14 ms	19.20 ms	42.06%
TodoMVC-JavaScript-ES6-Webpack-Complex-DOM	37.25 ms	19.45 ms	47.79%
TodoMVC-WebComponents	13.64 ms	8.76 ms	35.78%
TodoMVC-React-Complex-DOM	27.23 ms	17.41 ms	36.06%
TodoMVC-React-Redux	28.60 ms	17.57 ms	38.57%
TodoMVC-Backbone	20.75 ms	12.51 ms	39.71%
TodoMVC-Angular-Complex-DOM	25.02 ms	16.03 ms	35.93%
TodoMVC-Vue	18.01 ms	11.62 ms	35.48%
TodoMVC-jQuery	95.79 ms	61.52 ms	35.77%
TodoMVC-Preact-Complex-DOM	11.05 ms	7.02 ms	36.47%
TodoMVC-Svelte-Complex-DOM	9.45 ms	6.38 ms	32.49%
TodoMVC-Lit-Complex-DOM	12.53 ms	8.17 ms	34.80%
NewsSite-Next	65.56 ms	37.84 ms	42.29%
NewsSite-Nuxt	51.96 ms	31.13 ms	40.08%
Editor-CodeMirror	20.09 ms	12.93 ms	35.64%
Editor-TipTap	55.42 ms	36.06 ms	34.93%
Charts-observable-plot	38.99 ms	24.34 ms	37.57%
Charts-chartjs	36.75 ms	23.29 ms	36.63%
React-Stockcharts-SVG	63.34 ms	39.33 ms	37.91%
Perf-Dashboard	-	16.76 ms	-

测试项

M1Pro

M5Pro

提升

TodoMVC-JavaScript-ES5

33.14 ms

19.20 ms

42.06%

TodoMVC-JavaScript-ES6-Webpack-Complex-DOM

37.25 ms

19.45 ms

47.79%

TodoMVC-WebComponents

13.64 ms

8.76 ms

35.78%

TodoMVC-React-Complex-DOM

27.23 ms

17.41 ms

36.06%

TodoMVC-React-Redux

28.60 ms

17.57 ms

38.57%

TodoMVC-Backbone

20.75 ms

12.51 ms

39.71%

TodoMVC-Angular-Complex-DOM

25.02 ms

16.03 ms

35.93%

TodoMVC-Vue

18.01 ms

11.62 ms

35.48%

TodoMVC-jQuery

95.79 ms

61.52 ms

35.77%

TodoMVC-Preact-Complex-DOM

11.05 ms

7.02 ms

36.47%

TodoMVC-Svelte-Complex-DOM

9.45 ms

6.38 ms

32.49%

TodoMVC-Lit-Complex-DOM

12.53 ms

8.17 ms

34.80%

NewsSite-Next

65.56 ms

37.84 ms

42.29%

NewsSite-Nuxt

51.96 ms

31.13 ms

40.08%

Editor-CodeMirror

20.09 ms

12.93 ms

35.64%

Editor-TipTap

55.42 ms

36.06 ms

34.93%

Charts-observable-plot

38.99 ms

24.34 ms

37.57%

Charts-chartjs

36.75 ms

23.29 ms

36.63%

React-Stockcharts-SVG

63.34 ms

39.33 ms

37.91%

Perf-Dashboard

16.76 ms

浏览器中，M5 Pro 相比于 M1 Pro，平均提升 ≈ 38.6%

编译 llvm

源码来自 https://github.com/llvm/llvm-project.git 的 8c84b3c4e28a2756f52295f21ef67e97bae000a1 commit

依赖版本如下：

cmake 4.2.3
ninja 1.13.2

配置、编译过程如下方命令：

$ git clone https://github.com/llvm/llvm-project.git --depth=1
$ cd llvm-project
$ mkdir build && cd build
$ cmake -G Ninja ../llvm \
  -DCMAKE_BUILD_TYPE=Release \
  -DLLVM_ENABLE_PROJECTS="clang;clang-tools-extra;lld;mlir" \
  -DCMAKE_INSTALL_PREFIX=/usr/local/llvm
$ ninja -j$(sysctl -n hw.ncpu)

最终结果 (时间数据来自 zsh 的 Powerlevel10K 主题)：

M1 Pro：32m 26s
M5 Pro：13m 11s，提升 ≈ 59.35%

复合 maven 工程测试

这是我司的 Java 后端工程，包括但不限于 cli、SpringBoot 等微服务，测试命令使用 mvn -T 1C -U -B -DskipTests -Dmaven.test.skip=true clean verify

测试条件

$ mvn --version
Apache Maven 3.9.14 (996c630dbc656c76214ce58821dcc58be960875b)
Maven home: /opt/homebrew/Cellar/maven/3.9.14/libexec
Java version: 21.0.10, vendor: Homebrew, runtime: /opt/homebrew/Cellar/openjdk@21/21.0.10/libexec/openjdk.jdk/Contents/Home
Default locale: zh_US_#Hans, platform encoding: UTF-8
OS name: "mac os x", version: "26.3.1", arch: "aarch64", family: "mac"

测试结果：

M1 Pro：Total time: 12.938 s (Wall Clock)
M5 Pro：Total time: 5.792 s (Wall Clock)，提升 ≈ 55.23%

ANE 神经网络引擎

本次测试使用仓库：anemll-bench

此仓库对于 M1 Pro 芯片认为没必要跑额外的 dual model 测试

M1 Pro

  标准 ANE benchmark 结果：
  - llama_lm_head: 19.11ms, 54.99 GB/s
  - llama_lm_head_lut6, 8.52ms, 46.51 GB/s
  - DeepHermes_lm_head, 19.10ms, 55.01 GB/s

M5 Pro

  标准 ANE benchmark 结果：

  - llama_lm_head: 6.99 ms, 150.38 GB/s
  - llama_lm_head_lut6: 3.66 ms, 108.12 GB/s
  - DeepHermes_lm_head: 7.01 ms, 149.80 GB/s

  Dual-model benchmark 结果：

  - 单跑：llama_lm_head 6.99 ms / 150.24 GB/s，DeepHermes_lm_head 6.99 ms / 150.30 GB/s
  - 并行：14.05 ms / 74.77 GB/s 和 14.03 ms / 74.90 GB/s
  - 合并吞吐：149.54 GB/s
  - 效率：49.76%

游戏

这部分测试 M1 Pro 不参与

Java 原版 Minecraft

此处使用了 HMCL 启动器，版本 3.9.1，游戏组件版本如下：

Minecraft 1.21.11
Forge 61.1.3
Sodium 0.8.6

详细的统计信息：

基本性能

帧率：约 115 FPS
垂直同步：关闭
最大帧率限制：无限制
全屏：关闭
当前分辨率：当前系统分辨率 (系统设置 → 显示器 → 更多空间后窗口最大化)

渲染与模拟

渲染距离：32 区块
模拟距离：32 区块
实体渲染距离：100%
区块构建器：线程化

图形质量

生物群系过渡距离：15×15（最高）
平滑光照：开启
云：高品质
云层距离：128 区块
实体阴影：开启
粒子效果：全部
天气效果半径：10 方块
Mipmap 级别：4
各向异性过滤：4×
纹理过滤：无

调试信息（F3）

XYZ：88.856 / 63 / -53.657
Block：88 63 -54
Facing：West
Chunk：5 3 -4

内存使用

已使用内存：1465 MB / 6144 MB（约 23%）
已分配堆：2368 MB（约 30%）
内存分配速率：约 141 MB/s

黑神话：悟空性能测试工具 (运行在 CrossOver)

测试条件和结果如图：

blackmyth benchmark

总结

此部分由 LLM 撰写

从整篇测试看，M5 Pro 对 M1 Pro 并不是一次“小修小补”，而是一次比较完整的代际跃迁。即便本文里给 M5 Pro 额外叠了散热、后台进程和供电条件上的 Debuff，它依然在绝大多数项目里稳定拉开差距：Cinebench GPU 从 9789 提升到 40175，接近 4.1 倍；CPU 多核从 2634 提升到 7702，约 2.9 倍；单核也从 442 提升到 738，约 1.67 倍。这说明它不只是峰值更高，单线程响应和并行吞吐都上了一个明显的台阶。

如果工作负载以“干活”为主，升级收益是最直观的。ffmpeg 视频合成从 6 分 33 秒缩短到 2 分 45 秒，Redis 编译从 31 秒降到 14 秒，llvm 全量编译从 32 分 26 秒降到 13 分 11 秒，复合 Maven 工程也从 12.938 秒降到 5.792 秒，整体都落在约 55% 到 60% 的缩短区间。换句话说，M5 Pro 带来的不是零散的小优化，而是每次构建、每次转码都能感知到的时间回收。

本地 AI 相关任务的提升也很实在。LM Studio 下，gemma-3-4b 的生成速度从 49.03 提升到 96.79 Token/s，qwen3.5-9b 从 16.2 提升到 34.78 Token/s，TTFT 也分别从 0.83s 降到 0.24s、1.32s 降到 0.71s；ANE 测试里，llama_lm_head 延迟从 19.11 ms 降到 6.99 ms，吞吐从 54.99 GB/s 提高到 150.38 GB/s。对于本地模型推理、代码补全、agent 类工作流来说，这种提升会直接反映在“等输出”的时间上。

内存与存储部分虽然没有 GPU 跑分那么夸张，但同样属于全线抬升。STREAM 平均带宽提升约 59.2%，rchase 随机访问平均延迟从 143.922 ns 降到 119.43 ns，4K 随机混合读写从 72.7 MiB/s / 18.6k IOPS 提升到 124 MiB/s / 31.8k IOPS，读写平均延迟也有明显改善。这些指标未必像 Cinebench 那样醒目，但会实打实地体现在大型工程、索引、依赖解析和多任务切换的体验里。

相对更克制的一项是浏览器渲染。Speedometer 3.0 的平均提升约为 38.6%，已经足够让页面交互、前端开发和复杂后台页更顺滑，但它的变化确实没有编译、视频处理和本地 AI 这么夸张。所以如果只看结论：M1 Pro 到今天依然不是不能用，但只要日常负载里有频繁编译、视频转码、本地模型或更重的并行任务，M5 Pro 就已经不只是“锦上添花”，而是能持续兑现到每天工作流里的升级。

Licensed under Anti-996 License 1.0
本文链接：https://epliar.com/articles/m1-pro-and-m5-pro-benchmark/

前言

测试环境变量

开始测试

Cinebench

LM Studio (大模型 Token/s 测试)

google/gemma-3-4b

qwen/qwen3.5-9b

ffmpeg 视频合成

硬盘读写测试

编译 Redis

内存持续带宽测试

rchase 内存随机访问测试

硬盘随机读写测试

Web 渲染性能测试 (Speedometer3.0)

编译 llvm

复合 maven 工程测试

ANE 神经网络引擎

游戏

Java 原版 Minecraft

基本性能

渲染与模拟

图形质量

调试信息（F3）

内存使用

黑神话：悟空 性能测试工具 (运行在 CrossOver)

总结

黑神话：悟空性能测试工具 (运行在 CrossOver)