<?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[Strata - 3090 / R9700, 2133 ddr4 / 2633 ddr4]]></title><description><![CDATA[<p dir="auto">刚换了 32gbx2 2133mhz 去 32gbx8 2633mhz<br />
神奇的是 速度完全没有改变<br />
所以 还没买内存的朋友 可能不需要追求3200 mhz 毕竟价钱 差不少</p>
<p dir="auto">原本是64gb 不够玩strata 又找不到便宜内存<br />
刚好本地有人出2633 但是只有买10条才肯给折扣<br />
所以忍痛买了</p>
<p dir="auto">顺便放下 单3090, 双3090, 单r9700 strata的对比<br />
epyc7452</p>
<h2>Decode ( greedy, 1024 tokens, median of 3; range in brackets)</h2>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>Workload</th>
<th>2x 3090</th>
<th>1x 3090</th>
<th>R9700</th>
<th>1x 3090 vs 2x</th>
<th>R9700 vs 2x</th>
</tr>
</thead>
<tbody>
<tr>
<td>Chinese prose</td>
<td><strong>104.8</strong> t/s (96.4-108.3)</td>
<td>83.4 (76.2-83.5)</td>
<td>64.2 (59.8-65.6)</td>
<td>-20 %</td>
<td>-39 %</td>
</tr>
<tr>
<td>Chinese chat</td>
<td><strong>112.3</strong> (110.0-112.7)</td>
<td>83.6 (82.3-84.2)</td>
<td>70.6 (66.5-72.6)</td>
<td>-26 %</td>
<td>-37 %</td>
</tr>
<tr>
<td>English technical</td>
<td><strong>109.2</strong> (103.7-109.9)</td>
<td>83.1 (74.9-84.2)</td>
<td>70.6 (65.6-72.8)</td>
<td>-24 %</td>
<td>-35 %</td>
</tr>
<tr>
<td>Code</td>
<td><strong>159.7</strong> (141.5-161.1)</td>
<td>117.5 (93.3-118.1)</td>
<td>108.4 (94.3-109.2)</td>
<td>-26 %</td>
<td>-32 %</td>
</tr>
</tbody>
</table>
<h2>Prefill (cold prompts with a unique salt)</h2>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>Prompt</th>
<th>2x 3090</th>
<th>1x 3090</th>
<th>R9700</th>
</tr>
</thead>
<tbody>
<tr>
<td>19K</td>
<td><strong>2,489</strong> t/s (7.7 s)</td>
<td>1,912 (10.0 s)</td>
<td>1,039 (18.5 s)</td>
</tr>
<tr>
<td>43K</td>
<td><strong>3,137</strong> (13.7 s)</td>
<td>2,005 (21.5 s)</td>
<td>1,323 (32.5 s)</td>
</tr>
<tr>
<td>89K</td>
<td><strong>3,523</strong> (25.2 s)</td>
<td>2,030 (43.8 s)</td>
<td>1,156 (76.9 s)</td>
</tr>
<tr>
<td>123K</td>
<td><strong>3,564</strong> (34.4 s)</td>
<td>1,994 (61.4 s)</td>
<td>1,123 (109.0 s)</td>
</tr>
</tbody>
</table>
<h2>IQ3_S vs IQ3_XXS on 2x RTX 3090</h2>
<h3>Decode</h3>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>Workload</th>
<th>IQ3_XXS, 64 GB RAM (2026-10-07)</th>
<th>IQ3_XXS, 256 GB RAM</th>
<th>IQ3_S, 256 GB RAM</th>
<th>IQ3_S vs IQ3_XXS (same RAM)</th>
</tr>
</thead>
<tbody>
<tr>
<td>Chinese prose</td>
<td>109.9 t/s</td>
<td>105.2 (102.1-106.8)</td>
<td>104.8 (96.4-108.3)</td>
<td>-0.4 %</td>
</tr>
<tr>
<td>Chinese chat</td>
<td>112.2</td>
<td>115.8 (114.4-116.4)</td>
<td>112.3 (110.0-112.7)</td>
<td>-3.0 %</td>
</tr>
<tr>
<td>English technical</td>
<td>111.6</td>
<td>114.8 (114.5-115.6)</td>
<td>109.2 (103.7-109.9)</td>
<td>-4.9 %</td>
</tr>
<tr>
<td>Code</td>
<td>163.5</td>
<td>166.5 (153.9-167.2)</td>
<td>159.7 (141.5-161.1)</td>
<td>-4.1 %</td>
</tr>
</tbody>
</table>
<h3>Prefill</h3>
<table class="table table-bordered table-striped">
<thead>
<tr>
<th>Prompt</th>
<th>IQ3_XXS, 64 GB RAM</th>
<th>IQ3_XXS, 256 GB RAM</th>
<th>IQ3_S, 256 GB RAM</th>
<th>IQ3_S vs IQ3_XXS (same RAM)</th>
</tr>
</thead>
<tbody>
<tr>
<td>19K</td>
<td>2,571 t/s</td>
<td>2,563 (7.5 s)</td>
<td>2,489 (7.7 s)</td>
<td>-2.9 %</td>
</tr>
<tr>
<td>43K</td>
<td>3,183</td>
<td>3,236 (13.3 s)</td>
<td>3,137 (13.7 s)</td>
<td>-3.1 %</td>
</tr>
<tr>
<td>89K</td>
<td>3,585</td>
<td>3,649 (24.4 s)</td>
<td>3,523 (25.2 s)</td>
<td>-3.5 %</td>
</tr>
<tr>
<td>123K</td>
<td>3,647</td>
<td>3,686 (33.2 s)</td>
<td>3,564 (34.4 s)</td>
<td>-3.3 %</td>
</tr>
</tbody>
</table>
<p dir="auto">单张 3090 胜过 R9700:解码快 8-30%(代码场景差距最小),预填充约快 1.5-1.8 倍。<br />
同一张 R9700 上的对比: vLLM 跑稠密模型 Qwen3.8-27B(MXFP4,DFlash2),解码为 42 / 52 / 81 / 181 t/s,19K 时预填充约 2,500 t/s。Strata 的 125B 在中文上更快(64-71 对 42-52),在代码和提示词处理上更慢。<br />
IQ3_S 相比 IQ3_XXS 的代价很小:解码仅慢 0-5%,预填充约慢 3%,在 256K 以内召回表现相同,且其作者评定它与完整的 BF16 模型持平。它现在是默认选项(~/start_strata.sh)。<br />
内存升级没有改变 2x 3090 的速度(IQ3_XXS 与 64 GB 时的数据相差在 ±3% 以内)。</p>
<p dir="auto">所以我暂时的决定就是双3090 跑strata， 单张r9700 跑vllm</p>
]]></description><link>https://lcz.me/topic/2266</link><generator>RSS for Node</generator><lastBuildDate>Mon, 12 Oct 2026 12:25:16 GMT</lastBuildDate><atom:link href="https://lcz.me/topic/2266.rss" rel="self" type="application/rss+xml"/><pubDate>Sun, 11 Oct 2026 11:18:39 GMT</pubDate><ttl>60</ttl><item><title><![CDATA[Reply to Strata - 3090 / R9700, 2133 ddr4 / 2633 ddr4 on Mon, 12 Oct 2026 03:28:14 GMT]]></title><description><![CDATA[<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/xiaote" aria-label="Profile: Xiaote">@<bdi>Xiaote</bdi></a> 我之前只有双通道64gb</p>
]]></description><link>https://lcz.me/post/24159</link><guid isPermaLink="true">https://lcz.me/post/24159</guid><dc:creator><![CDATA[applejuice]]></dc:creator><pubDate>Mon, 12 Oct 2026 03:28:14 GMT</pubDate></item><item><title><![CDATA[Reply to Strata - 3090 / R9700, 2133 ddr4 / 2633 ddr4 on Mon, 12 Oct 2026 01:05:18 GMT]]></title><description><![CDATA[<p dir="auto">让AI调一下呢?有个calibrate脚本的 . 我让AI给我测了几组,最好的是kv resident 65536 tokens, 然后CPU使用32个物理核心(总共64核心)</p>
]]></description><link>https://lcz.me/post/24111</link><guid isPermaLink="true">https://lcz.me/post/24111</guid><dc:creator><![CDATA[stxpnet]]></dc:creator><pubDate>Mon, 12 Oct 2026 01:05:18 GMT</pubDate></item><item><title><![CDATA[Reply to Strata - 3090 / R9700, 2133 ddr4 / 2633 ddr4 on Mon, 12 Oct 2026 00:02:33 GMT]]></title><description><![CDATA[<p dir="auto">内存频率没带来提升，符合这个负载的特征，不是测错：</p>
<ul>
<li>decode 阶段 MoE 每个 token 只读激活专家，你这颗 125B 的激活参数不大，104 t/s 对应的实际带宽需求大约 100 GB/s 量级；DDR4-2133 八通道理论就有约 136 GB/s、2633 约 168 GB/s，两边都在需求之上，所以瓶颈不在系统内存带宽。</li>
<li>之前 64G 不够是<strong>容量</strong>问题（专家/PLE 放不下 → 换页），不是<strong>速度</strong>问题；换 256G 解决的是"装得下"，不是"跑得快"。你的数据也同向：64G → 256G 后 IQ3_XXS 速度反而略降（约 -3%），说明不带宽敏感。</li>
<li>所以 3200 对纯 decode 也不会明显加速，除非把 dense 权重/KV 也压到内存，或扩大 batch/上下文让激活集变大。优先"插满通道 + 够大"，2133 与 2633 的价差可以留给别处。</li>
<li>单卡 3090 对 R9700 的 prefill 快 1.5–1.8×，主轴是显存带宽（936 vs 640 GB/s）加 WMMA 路径，跟系统内存无关。</li>
</ul>
]]></description><link>https://lcz.me/post/24103</link><guid isPermaLink="true">https://lcz.me/post/24103</guid><dc:creator><![CDATA[Xiaote]]></dc:creator><pubDate>Mon, 12 Oct 2026 00:02:33 GMT</pubDate></item><item><title><![CDATA[Reply to Strata - 3090 / R9700, 2133 ddr4 / 2633 ddr4 on Sun, 11 Oct 2026 11:27:07 GMT]]></title><description><![CDATA[<p dir="auto">我买的是2666 的超频 3200 使用,</p>
]]></description><link>https://lcz.me/post/23993</link><guid isPermaLink="true">https://lcz.me/post/23993</guid><dc:creator><![CDATA[exllm]]></dc:creator><pubDate>Sun, 11 Oct 2026 11:27:07 GMT</pubDate></item><item><title><![CDATA[Reply to Strata - 3090 / R9700, 2133 ddr4 / 2633 ddr4 on Sun, 11 Oct 2026 11:24:35 GMT]]></title><description><![CDATA[<p dir="auto"><a class="plugin-mentions-user plugin-mentions-a" href="/user/terry" aria-label="Profile: terry">@<bdi>terry</bdi></a> 所以我也没买3200 只是顺便测试</p>
<p dir="auto">刚好本地的有人出2633<br />
价钱是跟我之前买2133 的时候差不多一样</p>
]]></description><link>https://lcz.me/post/23989</link><guid isPermaLink="true">https://lcz.me/post/23989</guid><dc:creator><![CDATA[applejuice]]></dc:creator><pubDate>Sun, 11 Oct 2026 11:24:35 GMT</pubDate></item><item><title><![CDATA[Reply to Strata - 3090 / R9700, 2133 ddr4 / 2633 ddr4 on Sun, 11 Oct 2026 11:20:53 GMT]]></title><description><![CDATA[<p dir="auto">不是做过了专门的视频吗，没啥影响<img src="https://lcz.me/assets/plugins/nodebb-plugin-emoji/emoji/android/1f622.png?v=fe25256ab4a" class="not-responsive emoji emoji-android emoji--cry" style="height:1.4em;width:auto;vertical-align:middle" title="😢" alt="😢" /></p>
]]></description><link>https://lcz.me/post/23988</link><guid isPermaLink="true">https://lcz.me/post/23988</guid><dc:creator><![CDATA[terry]]></dc:creator><pubDate>Sun, 11 Oct 2026 11:20:53 GMT</pubDate></item></channel></rss>