deepseek-v4.1IMAGE
4 台 DGX Spark 跑 V4.1 Flash:双线 RoCE 让长提示读取快 15%
模型大到单台 DGX Spark 装不下,被拆到 4 台上,每步经 RoCE 交换;加第二条线让通道翻倍,长提示读取速度提升约 15%。
Tech2Wild
@Tech2Wild12 天前收录原推发布于 2026.10.18

Tech2Wild @Tech2Wild原推发布于 2026.10.18
🧠 Why a 2nd cable = more speed: DeepSeek V4.1 Flash is too big for one DGX Spark, so it's split across 4. Every step, the Sparks swap their work over RoCE. 1 cable = 1 lane. 2 cables = double the lanes 🛣️ Long prompts move the most data, so reading them jumped +15% 📖
这条是引用着另一条写的:「双网线 DGX Spark 本地跑 V4.1 Flash:单用户 93.6 tok/s」
x.com/tech2wild/status/2111666660101026191采集于 2026.10.18 16:25