跳到正文
DEV Community· News·· 18 天前AI 評分64

SambaNova 與 Supermicro 合作,Supermicro 將採購 10,000 顆 SN50 RDU 用於分離式 AI 推理系統

SambaNova and Supermicro Partner on SN50 Chips for Disaggregated AI Inference

AI 導讀

SambaNova 與 Supermicro 宣佈圍繞 SN50 RDU 合作,Supermicro 將先採購 10,000 顆,用於結合 NVIDIA GPU 與 SambaNova RDU 的分離式推理系統。SN50 於 2026 年 2 月推出;廠商稱其速度是競品的 5 倍、TCO 低 3 倍,這些數字並非獨立測試結果。交付將從 neocloud 和主權雲服務商開始。

正文

News

Infographic cover

📅 4 November 2026 • datacenterdynamics.com

SambaNova and Supermicro Partner on SN50 Chips for Disaggregated AI Inference

SambaNova and Supermicro announced a partnership around the SN50 RDU, and the commitment is concrete: Supermicro will initially purchase 10,000 SN50 RDUs for disaggregated inference systems that combine NVIDIA GPUs with SambaNova RDUs. The SN50 was introduced in February 2026. According to company claims, the chip runs 5 times faster than competing chips and delivers 3 times lower TCO, vendor figures rather than independently measured results.

The architecture explains why two chip types share a rack. SN50 uses three tier memory holding models up to 10 trillion parameters, and up to 256 accelerators link through an interconnect with bandwidth in the tens of terabits per second. Disaggregated inference splits serving stages across suited hardware, and agentic workloads reward exactly that: long tool calling chains stress prefill compute and memory capacity differently from single prompt chat.

Delivery starts with neocloud and sovereign cloud providers, where inference demand is loudest. These buyers rent compute by the token and optimize throughput per dollar rather than benchmark records, so pairing RDU efficiency with GPU compatibility attacks the procurement decision directly. Supermicro's integration role matters as much as the chip, since rack level integration is where deployment friction lives. In practice, integration quality decides results as much as chip speed.

Impact. Inference hardware is specializing beyond GPUs, and hybrid racks are becoming a shipped configuration rather than an experiment. Treat vendor claims as claims until measured on your workloads, and watch neocloud rental pricing as the fastest public signal of real cost per dollar. If the numbers hold, the cost per token of agentic applications falls for every team paying for inference.

Sources: datacenterdynamics.com

Key figures

IMPACT checklist


Continue the conversation: https://t.me/AIAutomationArchitecture · Bluesky

AI Automation

來源:DEV Community · dev.to