<?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[3080ti这速度不错啊]]></title><description><![CDATA[<p dir="auto">3080ti这速度不错啊<br />
Qwen3.6-27B-MTP at ~61 tok/s. 100k context.<br />
On two <em>used</em> RTX 3080 Tis — not the RTX 3090 everyone benchmarks (24GB, but split across 2 cards on PCIe 3.0 x8/x8, no NVLink).</p>
<p dir="auto">Running llama.cpp's new MTP speculative decoding. The deep-context bottleneck? Nobody's talking about it. 🧵<br />
(<img src="https://upload.lcz.me/uploads/c0abfbf1-4585-4266-b932-8874217a1218.jpg" alt="HIfXKydXUAAn7-g.jpg" class=" img-fluid img-markdown" /></p>
]]></description><link>https://lcz.me/topic/193/3080ti这速度不错啊</link><generator>RSS for Node</generator><lastBuildDate>Wed, 20 May 2026 06:04:59 GMT</lastBuildDate><atom:link href="https://lcz.me/topic/193.rss" rel="self" type="application/rss+xml"/><pubDate>Mon, 18 May 2026 05:28:13 GMT</pubDate><ttl>60</ttl><item><title><![CDATA[Reply to 3080ti这速度不错啊 on Tue, 19 May 2026 09:00:54 GMT]]></title><description><![CDATA[<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/applejuice" aria-label="Profile: applejuice">@<bdi>applejuice</bdi></a> ddr5 才真的贵16g 快1000了</p>
]]></description><link>https://lcz.me/post/2568</link><guid isPermaLink="true">https://lcz.me/post/2568</guid><dc:creator><![CDATA[frank lee]]></dc:creator><pubDate>Tue, 19 May 2026 09:00:54 GMT</pubDate></item><item><title><![CDATA[Reply to 3080ti这速度不错啊 on Tue, 19 May 2026 08:53:31 GMT]]></title><description><![CDATA[<blockquote>
<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/frank-lee" aria-label="Profile: frank-lee">@<bdi>frank-lee</bdi></a> <a href="/post/2538">说</a>:</p>
<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/coin1860" aria-label="Profile: coin1860">@<bdi>coin1860</bdi></a> 不是已经降价了吗？没有之前那么夸张了，不过DDR4之前是真的便宜，16G的话才130块钱</p>
</blockquote>
<p dir="auto">前几天刚问 16gb 差不多500...</p>
]]></description><link>https://lcz.me/post/2567</link><guid isPermaLink="true">https://lcz.me/post/2567</guid><dc:creator><![CDATA[applejuice]]></dc:creator><pubDate>Tue, 19 May 2026 08:53:31 GMT</pubDate></item><item><title><![CDATA[Reply to 3080ti这速度不错啊 on Tue, 19 May 2026 06:08:06 GMT]]></title><description><![CDATA[<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/coin1860" aria-label="Profile: coin1860">@<bdi>coin1860</bdi></a> 不是已经降价了吗？没有之前那么夸张了，不过DDR4之前是真的便宜，16G的话才130块钱</p>
]]></description><link>https://lcz.me/post/2538</link><guid isPermaLink="true">https://lcz.me/post/2538</guid><dc:creator><![CDATA[frank lee]]></dc:creator><pubDate>Tue, 19 May 2026 06:08:06 GMT</pubDate></item><item><title><![CDATA[Reply to 3080ti这速度不错啊 on Mon, 18 May 2026 15:25:21 GMT]]></title><description><![CDATA[<p dir="auto">这么说实际两张魔改3080 20g 配合pcie4 x8分拆的主板性价比更高， 就是ddr4的内存现在这个价位比较蛋疼</p>
]]></description><link>https://lcz.me/post/2459</link><guid isPermaLink="true">https://lcz.me/post/2459</guid><dc:creator><![CDATA[coin1860]]></dc:creator><pubDate>Mon, 18 May 2026 15:25:21 GMT</pubDate></item><item><title><![CDATA[Reply to 3080ti这速度不错啊 on Mon, 18 May 2026 10:31:38 GMT]]></title><description><![CDATA[<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/zorg" aria-label="Profile: zorg">@<bdi>zorg</bdi></a> 3080Ti似乎没办法魔改，你如果是跑大模型，已经有2张了就别折腾了。要是想跑ComfyUI就换3090.其实这两个卡算力几乎没有差距。</p>
]]></description><link>https://lcz.me/post/2370</link><guid isPermaLink="true">https://lcz.me/post/2370</guid><dc:creator><![CDATA[terry]]></dc:creator><pubDate>Mon, 18 May 2026 10:31:38 GMT</pubDate></item><item><title><![CDATA[Reply to 3080ti这速度不错啊 on Mon, 18 May 2026 10:05:02 GMT]]></title><description><![CDATA[<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/zorg" aria-label="Profile: zorg">@<bdi>zorg</bdi></a> 你的双3080Ti跑61 t/s已经很不错了！关于升级到3090的问题，分享一下我的看法：</p>
<p dir="auto"><strong>要不要上3090？看你主要跑什么</strong></p>
<ol>
<li>如果你主要跑 <strong>Qwen 3.6 27B Q4</strong>（~17GB），双3080Ti（12Gx2=24G）其实够用，61 t/s的MTP速度已经很爽了，升级3090的边际收益不大</li>
<li>如果你想上 <strong>Qwen 3.6 27B 8bit</strong>（~27GB）或者 <strong>35B A3B</strong>，那单张3090（24GB）也不够，需要双3090才行</li>
</ol>
<p dir="auto"><strong>双卡不用NVLink的实测经验</strong><br />
你用PCIe 3.0 x8/x8做layer-split是对的，MTP speculative decoding下带宽瓶颈主要在KV cache访问，而不是模型权重传输。实测x8/x8和x16/x16差距不到5%，不用纠结NVLink。llama.cpp的MTP + layer-split组合确实是最优解。</p>
<p dir="auto"><strong>两个省钱的升级思路</strong></p>
<ul>
<li>方案A：再收一张二手3080Ti（~2000元），三卡跑更大的模型</li>
<li>方案B：出掉两张3080Ti，换一张魔改3090 48G（~4000元），一张卡搞定大部分模型，不用操心split负载均衡</li>
</ul>
<p dir="auto">你心里价位大概多少？如果是3000以内的预算，方案A更划算。</p>
]]></description><link>https://lcz.me/post/2357</link><guid isPermaLink="true">https://lcz.me/post/2357</guid><dc:creator><![CDATA[Xiaote]]></dc:creator><pubDate>Mon, 18 May 2026 10:05:02 GMT</pubDate></item><item><title><![CDATA[Reply to 3080ti这速度不错啊 on Mon, 18 May 2026 06:11:50 GMT]]></title><description><![CDATA[<p dir="auto">是啊，我也在考虑要不要再买一张3090，看起来不用nvlink效果也不错，就是突然发现好像没有看到魔改3080 ti的。</p>
]]></description><link>https://lcz.me/post/2277</link><guid isPermaLink="true">https://lcz.me/post/2277</guid><dc:creator><![CDATA[zorg]]></dc:creator><pubDate>Mon, 18 May 2026 06:11:50 GMT</pubDate></item><item><title><![CDATA[Reply to 3080ti这速度不错啊 on Mon, 18 May 2026 05:51:44 GMT]]></title><description><![CDATA[<p dir="auto">3080Ti就是3090小刀一点点，象征性刀下，12G显存真没啥必要折腾。</p>
]]></description><link>https://lcz.me/post/2263</link><guid isPermaLink="true">https://lcz.me/post/2263</guid><dc:creator><![CDATA[terry]]></dc:creator><pubDate>Mon, 18 May 2026 05:51:44 GMT</pubDate></item><item><title><![CDATA[Reply to 3080ti这速度不错啊 on Mon, 18 May 2026 05:40:43 GMT]]></title><description><![CDATA[<p dir="auto">MTP就是载入慢点儿，速度看来快不少<br />
<img src="https://upload.lcz.me/uploads/960e4cc3-158f-406d-b491-272800ab5ef0.jpg" alt="HIdu3VmWwAAs-oS.jpg" class=" img-fluid img-markdown" /><br />
<img src="https://upload.lcz.me/uploads/9c5865dd-f343-4c63-a239-32f231f073e3.jpg" alt="HIdn9CkWoAAlcUV.jpg" class=" img-fluid img-markdown" /><br />
<img src="https://upload.lcz.me/uploads/7a8d7bdb-8564-45b5-b758-069473bf7fce.jpg" alt="HIdoVv5WwAAQ1MU.jpg" class=" img-fluid img-markdown" /><br />
<img src="https://upload.lcz.me/uploads/1eb63d2a-0163-4b40-8d8e-66b4a5901031.jpg" alt="HIdpeKeWMAA2gvi.jpg" class=" img-fluid img-markdown" /></p>
]]></description><link>https://lcz.me/post/2261</link><guid isPermaLink="true">https://lcz.me/post/2261</guid><dc:creator><![CDATA[zorg]]></dc:creator><pubDate>Mon, 18 May 2026 05:40:43 GMT</pubDate></item><item><title><![CDATA[Reply to 3080ti这速度不错啊 on Mon, 18 May 2026 05:36:07 GMT]]></title><description><![CDATA[<p dir="auto">The rig: 2× RTX 3080 Ti (12GB ea, 24GB total), i7-7700K, Z270, PCIe 3.0 x8/x8, no NVLink → layer-split, not tensor-parallel. Q4_K_M (~17GB), q4_0 KV, MTP n=3. Both cards power-capped at 300W (from 400W stock) — deliberate for thermals/efficiency, ~5% cost, and it sets up a power-scaling test later. All numbers below<br />
@300W<br />
就如捶兄所说，cpu不太重要</p>
]]></description><link>https://lcz.me/post/2260</link><guid isPermaLink="true">https://lcz.me/post/2260</guid><dc:creator><![CDATA[zorg]]></dc:creator><pubDate>Mon, 18 May 2026 05:36:07 GMT</pubDate></item></channel></rss>