<?xml version="1.0" encoding="UTF-8"?><rss xmlns:dc="http://purl.org/dc/elements/1.1/" xmlns:content="http://purl.org/rss/1.0/modules/content/" xmlns:atom="http://www.w3.org/2005/Atom" version="2.0"><channel><title><![CDATA[M5 Ultra 256G媒体机实测数据整理抢先看]]></title><description><![CDATA[<h1>M5 Ultra 256 G Mac Studio 上手实测：27B 本地模型 + 全套跑分（附 M3 Ultra 对比）</h1>
<blockquote>
<p dir="auto">汇总自一段 M5 Ultra 256G上手体验视频的实测视频。所有数字均为视频中展示的原始数据，未做二次加工。<br />
对比机型统一为 <strong>M3 Ultra</strong>。</p>
</blockquote>
<hr />
<h2>一、测试环境与口径</h2>
<p dir="auto"><strong>本地大模型测试配置</strong></p>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>项目</th>
<th>设置</th>
</tr>
</thead>
<tbody>
<tr>
<td>模型</td>
<td><strong>Qwen3.8-27B-4bit</strong>（15.7 GB）</td>
</tr>
<tr>
<td>Benchmark Context</td>
<td>Code (Python)</td>
</tr>
<tr>
<td>单请求测试项</td>
<td>pp1024 / pp4096 / pp8192</td>
</tr>
<tr>
<td>生成长度</td>
<td><strong>128 tokens（固定）</strong></td>
</tr>
<tr>
<td>ANE-aligned prompts</td>
<td><img src="https://lcz.me/assets/plugins/nodebb-plugin-emoji/emoji/android/2705.png?v=ecb7c61779a" class="not-responsive emoji emoji-android emoji--white_check_mark" style="height:23px;width:auto;vertical-align:middle" title="✅" alt="✅" /> 开启（+1 token，使 prefill 对齐 2048 整块）</td>
</tr>
<tr>
<td>连续批处理测试</td>
<td>2x / 4x / 8x（口径 pp1024 / tg128）</td>
</tr>
<tr>
<td>Force mix-lm engine</td>
<td><img src="https://lcz.me/assets/plugins/nodebb-plugin-emoji/emoji/android/2705.png?v=ecb7c61779a" class="not-responsive emoji emoji-android emoji--white_check_mark" style="height:23px;width:auto;vertical-align:middle" title="✅" alt="✅" /> 开启 → <strong>结果为 LM-only 基线（MTP 关闭）</strong></td>
</tr>
<tr>
<td>外部 API 端点</td>
<td>未启用（纯本地推理）</td>
</tr>
</tbody>
</table>
<blockquote>
<p dir="auto"><img src="https://lcz.me/assets/plugins/nodebb-plugin-emoji/emoji/android/26a0.png?v=ecb7c61779a" class="not-responsive emoji emoji-android emoji--warning" style="height:23px;width:auto;vertical-align:middle" title="⚠" alt="⚠" />️ 由于 MTP 被关闭，以下推理数字属于<strong>保守基线</strong>，不是该机日常可用速度。<br />
另外因开启了 ANE 对齐，测试项 pp1024/pp4096 在结果表中显示为 <strong>pp1025 / pp4097</strong>。</p>
</blockquote>
<hr />
<h2>二、AI 本地推理</h2>
<h3>2.1 单请求原始结果</h3>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>Test</th>
<th>TTFT</th>
<th>TPOT</th>
<th>pp TPS</th>
<th>tg TPS</th>
<th>E2E Latency</th>
<th>Throughput</th>
<th>Peak Mem</th>
</tr>
</thead>
<tbody>
<tr>
<td>pp1025/tg128</td>
<td>661.8 ms</td>
<td>20.84 ms/tok</td>
<td><strong>1548.7</strong> tok/s</td>
<td><strong>48.4</strong> tok/s</td>
<td>3.318 s</td>
<td>347.5 tok/s</td>
<td>15.85 GB</td>
</tr>
<tr>
<td>pp4097/tg128</td>
<td>2573.6 ms</td>
<td>21.09 ms/tok</td>
<td><strong>1591.9</strong> tok/s</td>
<td><strong>47.8</strong> tok/s</td>
<td>5.265 s</td>
<td>802.5 tok/s</td>
<td>17.01 GB</td>
</tr>
</tbody>
</table>
<h3>2.2 解码速度（单路）</h3>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>机型</th>
<th>解码速度</th>
</tr>
</thead>
<tbody>
<tr>
<td><strong>M5 Ultra</strong></td>
<td><strong>51</strong> Token/s</td>
</tr>
<tr>
<td>M3 Ultra</td>
<td>35 Token/s</td>
</tr>
</tbody>
</table>
<h3>2.3 多并发解码（聚合吞吐）</h3>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>并发数</th>
<th>M5 Ultra</th>
<th>M3 Ultra</th>
<th>画面标注</th>
</tr>
</thead>
<tbody>
<tr>
<td>2 并发</td>
<td><strong>102.4</strong></td>
<td>70.0</td>
<td><strong>+46.3%</strong></td>
</tr>
<tr>
<td>4 并发</td>
<td><strong>169.4</strong></td>
<td>130.0</td>
<td><strong>+30.3%</strong></td>
</tr>
<tr>
<td><strong>8 并发</strong></td>
<td>248.9</td>
<td><strong>257.4</strong></td>
<td><strong>−3.3%</strong> <img src="https://lcz.me/assets/plugins/nodebb-plugin-emoji/emoji/android/26a0.png?v=ecb7c61779a" class="not-responsive emoji emoji-android emoji--warning" style="height:23px;width:auto;vertical-align:middle" title="⚠" alt="⚠" />️</td>
</tr>
</tbody>
</table>
<blockquote>
<p dir="auto">8 并发是本组中唯一一项 M5 Ultra 落后 M3 Ultra 的测试。</p>
</blockquote>
<h3>2.4 多并发预填充</h3>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>并发数</th>
<th>M5 Ultra</th>
<th>M3 Ultra</th>
<th>画面标注</th>
</tr>
</thead>
<tbody>
<tr>
<td>2 并发</td>
<td><strong>1243.3</strong></td>
<td>345.5</td>
<td><strong>3.6×</strong></td>
</tr>
<tr>
<td>4 并发</td>
<td><strong>1071.9</strong></td>
<td>302.4</td>
<td><strong>3.5×</strong></td>
</tr>
<tr>
<td>8 并发</td>
<td><strong>1016.1</strong></td>
<td>282.6</td>
<td><strong>3.6×</strong></td>
</tr>
</tbody>
</table>
<blockquote>
<p dir="auto">横轴为并发数；纵轴画面标注为「每秒生成 token」。</p>
</blockquote>
<h3>2.5 8 并发端到端延迟</h3>
<p dir="auto">口径 <strong>pp1024 / tg512</strong> ｜ 单位：秒，越低越好</p>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>机型</th>
<th>端到端延迟</th>
</tr>
</thead>
<tbody>
<tr>
<td><strong>M5 Ultra</strong></td>
<td><strong>12.176 s</strong></td>
</tr>
<tr>
<td>M3 Ultra</td>
<td>32.971 s</td>
</tr>
</tbody>
</table>
<p dir="auto">画面标注：<strong>快 2.8 倍</strong></p>
<hr />
<h2>三、CPU</h2>
<h3>3.1 Geekbench 7 CPU</h3>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>项目</th>
<th>M5 Ultra</th>
<th>M3 Ultra</th>
<th>提升</th>
</tr>
</thead>
<tbody>
<tr>
<td>单核</td>
<td><strong>3,711</strong></td>
<td>2,963</td>
<td><strong>+25.2%</strong></td>
</tr>
<tr>
<td>多核</td>
<td><strong>51,760</strong></td>
<td>40,271</td>
<td><strong>+28.5%</strong></td>
</tr>
</tbody>
</table>
<h3>3.2 Cinebench 2026 CPU · 单核</h3>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>机型</th>
<th>分数</th>
</tr>
</thead>
<tbody>
<tr>
<td>锐龙 9950X3D</td>
<td>780</td>
</tr>
<tr>
<td><strong>M5 Ultra</strong></td>
<td><strong>744</strong></td>
</tr>
<tr>
<td>M3 Ultra</td>
<td>676</td>
</tr>
</tbody>
</table>
<h3>3.3 Cinebench 2026 CPU · 多核</h3>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>机型</th>
<th>分数</th>
<th>提升</th>
</tr>
</thead>
<tbody>
<tr>
<td><strong>M5 Ultra</strong></td>
<td><strong>17,348</strong></td>
<td>—</td>
</tr>
<tr>
<td>M3 Ultra</td>
<td>12,082</td>
<td><strong>+43.6%</strong></td>
</tr>
</tbody>
</table>
<hr />
<h2>四、GPU 合成基准</h2>
<h3>4.1 Geekbench 7 GPU</h3>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>机型</th>
<th>分数</th>
<th>提升</th>
</tr>
</thead>
<tbody>
<tr>
<td><strong>M5 Ultra</strong></td>
<td><strong>352,444</strong></td>
<td>—</td>
</tr>
<tr>
<td>M3 Ultra</td>
<td>263,242</td>
<td><strong>+33.9%</strong></td>
</tr>
</tbody>
</table>
<h3>4.2 Cinebench 2026 GPU</h3>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>机型</th>
<th>分数</th>
</tr>
</thead>
<tbody>
<tr>
<td>RTX 5090</td>
<td>160,000</td>
</tr>
<tr>
<td><strong>M5 Ultra</strong></td>
<td><strong>140,734</strong></td>
</tr>
<tr>
<td>RTX 5080</td>
<td>100,000</td>
</tr>
<tr>
<td>M3 Ultra</td>
<td>83,865</td>
</tr>
</tbody>
</table>
<p dir="auto">M5 Ultra 对 M3 Ultra：<strong>+67.8%</strong></p>
<h3>4.3 Blender GPU 测试</h3>
<p dir="auto">版本 <strong>5.2.1 LTS</strong> ｜ 单位：采样/分钟，越高越好</p>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>场景</th>
<th>M5 Ultra</th>
<th>M3 Ultra</th>
<th>提升</th>
</tr>
</thead>
<tbody>
<tr>
<td>杂货店</td>
<td><strong>3,795</strong></td>
<td>2,048</td>
<td><strong>+85.3%</strong></td>
</tr>
<tr>
<td>狮子</td>
<td><strong>4,806</strong></td>
<td>2,861</td>
<td><strong>+68%</strong></td>
</tr>
</tbody>
</table>
<h3>4.4 3DMark Steel Nomad</h3>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>机型</th>
<th>分数</th>
</tr>
</thead>
<tbody>
<tr>
<td>RTX 5080</td>
<td>8,800</td>
</tr>
<tr>
<td><strong>M5 Ultra</strong></td>
<td><strong>8,052</strong></td>
</tr>
<tr>
<td>M3 Ultra</td>
<td>5,548</td>
</tr>
<tr>
<td>RTX 5070</td>
<td>5,326</td>
</tr>
</tbody>
</table>
<p dir="auto">M5 Ultra 对 M3 Ultra：<strong>+45.1%</strong></p>
<h3>4.5 3DMark Solar Bay Extreme</h3>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>机型</th>
<th>分数</th>
</tr>
</thead>
<tbody>
<tr>
<td>RTX 5070 Ti</td>
<td>27,915</td>
</tr>
<tr>
<td><strong>M5 Ultra</strong></td>
<td><strong>26,105</strong></td>
</tr>
<tr>
<td>M3 Ultra</td>
<td>16,426</td>
</tr>
<tr>
<td>RTX 5060 Ti</td>
<td>15,300</td>
</tr>
</tbody>
</table>
<p dir="auto">M5 Ultra 对 M3 Ultra：<strong>+58.9%</strong></p>
<hr />
<h2>五、光追与游戏实机</h2>
<h3>赛博朋克 2077 基准测试</h3>
<p dir="auto">口径：<strong>4K 分辨率 · 超级光追预设 · MetalFX 质量档</strong> ｜ 单位：每秒帧数，越高越好</p>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>项目</th>
<th>M5 Ultra</th>
<th>M3 Ultra</th>
<th>提升</th>
</tr>
</thead>
<tbody>
<tr>
<td><strong>开启光追</strong></td>
<td><strong>37.28</strong></td>
<td>20.39</td>
<td><strong>+82.8%</strong></td>
</tr>
<tr>
<td><strong>关闭光追</strong></td>
<td><strong>60.32</strong></td>
<td>47.78</td>
<td><strong>+26.2%</strong></td>
</tr>
</tbody>
</table>
<hr />
<h2>六、提升幅度一览（M5 Ultra 对 M3 Ultra）</h2>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>测试项</th>
<th>提升</th>
</tr>
</thead>
<tbody>
<tr>
<td>Blender · 杂货店</td>
<td><strong>+85.3%</strong></td>
</tr>
<tr>
<td>赛博朋克 2077 · 开光追</td>
<td><strong>+82.8%</strong></td>
</tr>
<tr>
<td>Blender · 狮子</td>
<td><strong>+68%</strong></td>
</tr>
<tr>
<td>Cinebench 2026 GPU</td>
<td><strong>+67.8%</strong></td>
</tr>
<tr>
<td>3DMark Solar Bay Extreme</td>
<td><strong>+58.9%</strong></td>
</tr>
<tr>
<td>多并发解码 · 2 并发</td>
<td><strong>+46.3%</strong></td>
</tr>
<tr>
<td>3DMark Steel Nomad</td>
<td><strong>+45.1%</strong></td>
</tr>
<tr>
<td>Cinebench 2026 CPU · 多核</td>
<td><strong>+43.6%</strong></td>
</tr>
<tr>
<td>Geekbench 7 GPU</td>
<td><strong>+33.9%</strong></td>
</tr>
<tr>
<td>多并发解码 · 4 并发</td>
<td><strong>+30.3%</strong></td>
</tr>
<tr>
<td>Geekbench 7 CPU · 多核</td>
<td><strong>+28.5%</strong></td>
</tr>
<tr>
<td>赛博朋克 2077 · 关光追</td>
<td><strong>+26.2%</strong></td>
</tr>
<tr>
<td>Geekbench 7 CPU · 单核</td>
<td><strong>+25.2%</strong></td>
</tr>
<tr>
<td>多并发解码 · 8 并发</td>
<td><strong>−3.3%</strong></td>
</tr>
<tr>
<td>8 并发端到端延迟</td>
<td><strong>快 2.8 倍</strong></td>
</tr>
<tr>
<td>多并发预填充</td>
<td><strong>3.5–3.6×</strong></td>
</tr>
</tbody>
</table>
<hr />
<h2>数据说明</h2>
<ol>
<li>本文所有数值均直接取自于视频内容，未做换算或修正，仅供参考。</li>
</ol>
]]></description><link>https://lcz.me/topic/1874</link><generator>RSS for Node</generator><lastBuildDate>Fri, 25 Sep 2026 20:36:00 GMT</lastBuildDate><atom:link href="https://lcz.me/topic/1874.rss" rel="self" type="application/rss+xml"/><pubDate>Tue, 22 Sep 2026 00:57:37 GMT</pubDate><ttl>60</ttl><item><title><![CDATA[Reply to M5 Ultra 256G媒体机实测数据整理抢先看 on Tue, 22 Sep 2026 09:03:49 GMT]]></title><description><![CDATA[<p dir="auto">相同的顶级设计，相同的台积电工艺。在合理的舞台上</p>
<p dir="auto">我比较你功耗大，我就理论上比你强。 没有一点电是白费的，不会违背物理规律。</p>
]]></description><link>https://lcz.me/post/20047</link><guid isPermaLink="true">https://lcz.me/post/20047</guid><dc:creator><![CDATA[张光璞]]></dc:creator><pubDate>Tue, 22 Sep 2026 09:03:49 GMT</pubDate></item><item><title><![CDATA[Reply to M5 Ultra 256G媒体机实测数据整理抢先看 on Tue, 22 Sep 2026 08:52:52 GMT]]></title><description><![CDATA[<blockquote>
<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/david-chen" aria-label="Profile: David-Chen">@<bdi>David-Chen</bdi></a> <a href="/post/19878">说</a>:</p>
<p dir="auto">我昨天也追了一些，老實說，一個這麼大記憶容量，拿QWEN3.8 27B測試，你說用BF16就算了，還是用量化的，感覺好好的一個機器被糟蹋了，還是說跑Deepseek或是其他128B以上開源模型會難看到不行？醜媳婦總是要見公婆，沒看到直球對決，真的很可惜，再等等其他人的測試</p>
</blockquote>
<p dir="auto">好几个都是27B，我觉得是官方要求的。</p>
<p dir="auto">一堆人在堆 ，这年头苹果也要做推广 ，纯果粉自来水都不够 用了咩。</p>
]]></description><link>https://lcz.me/post/20044</link><guid isPermaLink="true">https://lcz.me/post/20044</guid><dc:creator><![CDATA[张光璞]]></dc:creator><pubDate>Tue, 22 Sep 2026 08:52:52 GMT</pubDate></item><item><title><![CDATA[Reply to M5 Ultra 256G媒体机实测数据整理抢先看 on Tue, 22 Sep 2026 07:48:54 GMT]]></title><description><![CDATA[<blockquote>
<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/cc-tgo" aria-label="Profile: cc-tgo">@<bdi>cc-tgo</bdi></a> <a href="/post/19874">说</a>:</p>
<p dir="auto">Qwen3.8-27B-4bit（15.7 GB）测试这个模型真的是毫无意义。</p>
</blockquote>
]]></description><link>https://lcz.me/post/20019</link><guid isPermaLink="true">https://lcz.me/post/20019</guid><dc:creator><![CDATA[Geekyang]]></dc:creator><pubDate>Tue, 22 Sep 2026 07:48:54 GMT</pubDate></item><item><title><![CDATA[Reply to M5 Ultra 256G媒体机实测数据整理抢先看 on Tue, 22 Sep 2026 07:37:13 GMT]]></title><description><![CDATA[<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/benton-yi" aria-label="Profile: benton-yi">@<bdi>benton-yi</bdi></a> 同感，m5u至少得是qwen3.8 next Flash起步才有意思</p>
]]></description><link>https://lcz.me/post/20016</link><guid isPermaLink="true">https://lcz.me/post/20016</guid><dc:creator><![CDATA[rock shi]]></dc:creator><pubDate>Tue, 22 Sep 2026 07:37:13 GMT</pubDate></item><item><title><![CDATA[Reply to M5 Ultra 256G媒体机实测数据整理抢先看 on Tue, 22 Sep 2026 04:01:31 GMT]]></title><description><![CDATA[<blockquote>
<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/david-chen" aria-label="Profile: David-Chen">@<bdi>David-Chen</bdi></a> <a href="/post/19878">说</a>:</p>
<p dir="auto">你說用BF16就算了，還是用量化的，感覺好好的一個機器被糟蹋了</p>
</blockquote>
<p dir="auto">如果成绩好的话他会不用嘛，27b稠密bf16别说m5u了，连pro6000成绩都没法看。苹果选测试对象非常有讲究的。<br />
小盒子2年内赌gb10阵营基本上问题不大。</p>
]]></description><link>https://lcz.me/post/19956</link><guid isPermaLink="true">https://lcz.me/post/19956</guid><dc:creator><![CDATA[benton yi]]></dc:creator><pubDate>Tue, 22 Sep 2026 04:01:31 GMT</pubDate></item><item><title><![CDATA[Reply to M5 Ultra 256G媒体机实测数据整理抢先看 on Tue, 22 Sep 2026 01:22:37 GMT]]></title><description><![CDATA[<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/david-chen" aria-label="Profile: David-Chen">@<bdi>David-Chen</bdi></a> 23日后应该会有一批用户实测出来吧，这个视频发出来后还被博主自己删除过，过了几小时才重发的。</p>
]]></description><link>https://lcz.me/post/19885</link><guid isPermaLink="true">https://lcz.me/post/19885</guid><dc:creator><![CDATA[cc tgo]]></dc:creator><pubDate>Tue, 22 Sep 2026 01:22:37 GMT</pubDate></item><item><title><![CDATA[Reply to M5 Ultra 256G媒体机实测数据整理抢先看 on Tue, 22 Sep 2026 01:07:10 GMT]]></title><description><![CDATA[<p dir="auto">我昨天也追了一些，老實說，一個這麼大記憶容量，拿QWEN3.8 27B測試，你說用BF16就算了，還是用量化的，感覺好好的一個機器被糟蹋了，還是說跑Deepseek或是其他128B以上開源模型會難看到不行？醜媳婦總是要見公婆，沒看到直球對決，真的很可惜，再等等其他人的測試</p>
]]></description><link>https://lcz.me/post/19878</link><guid isPermaLink="true">https://lcz.me/post/19878</guid><dc:creator><![CDATA[David Chen]]></dc:creator><pubDate>Tue, 22 Sep 2026 01:07:10 GMT</pubDate></item><item><title><![CDATA[Reply to M5 Ultra 256G媒体机实测数据整理抢先看 on Tue, 22 Sep 2026 01:02:44 GMT]]></title><description><![CDATA[<p dir="auto">补几个读数口径，方便和 x86 / N 卡横向比：</p>
<ol>
<li>
<p dir="auto">15.7 GB 权重、48.4 t/s 反推 decode 有效带宽约 760 GB/s（51 t/s 档约 800 GB/s），基本贴着这代统一内存的实际可达上限——单路 speed 已到顶，再往上主要靠 MTP / 连续批处理。</p>
</li>
<li>
<p dir="auto">表里 M5 51 t/s vs M3 35 t/s 不要只记绝对数，要连引擎设置一起看（是否同为 MTP 关、同为 mix-lm 基线），否则会把架构差异和配置差异混在一起。</p>
</li>
<li>
<p dir="auto">多并发建议同时给「聚合 t/s」和「单路 t/s」，不然 8x 的单路掉速会被聚合吞吐盖掉。</p>
</li>
</ol>
<p dir="auto">这组是关了 MTP 的保守基线，日常可用速度应该更高一些。</p>
]]></description><link>https://lcz.me/post/19876</link><guid isPermaLink="true">https://lcz.me/post/19876</guid><dc:creator><![CDATA[Xiaote]]></dc:creator><pubDate>Tue, 22 Sep 2026 01:02:44 GMT</pubDate></item></channel></rss>