What did Chengdu-based XFEON unveil in its September 2026 physical-AI product matrix, and how do the Xinghe S1 chip’s architecture, performaIllustration of XFEON’s proposed physical-AI stack; not a photograph of its products.
AI 提示
Create a landscape editorial hero image for this Studio Global article: What did Chengdu-based XFEON unveil in its September 2026 physical-AI product matrix, and how do the Xinghe S1 chip’s architecture, performa. Article summary: XFEON presented its September 2026 matrix as a proposed loop from **edge chips to physical action**: run models locally, collect real-world data, improve the models, and deploy them back to devices. The available reporti. Topic tags: general, general web, user generated. Style: premium digital editorial illustration, source-backed research mood, clean composition, high detail, modern web publication hero. Use reference image context only for broad subject, composition, and topical grounding; do not copy the exact image. Avoid: logos, brand marks, copyrighted characters, real person likenesses, fake screenshots, UI text, readable text, watermarks, charts with fa
openai.com
星凡智能(XFEON)總部位於成都。它在 2026 年 9 月 22 日公布的物理 AI 產品矩陣,想串起一個循環:讓 AI 模型在設備本地運行,從真實環境收集數據,再用數據改進模型,最後把模型用於決策和行動。這是公司的產品部署方向,不是整套系統已達到特定效能的證明。645
星核 S1:針對端側限制而設
星核 S1 瞄準功耗、散熱空間及記憶體都有限的物理 AI 設備。現有介紹指,它採用硬件支援的混合精度運算、低位元推理、稀疏化優化,以及晶片與演算法協同設計。配套的 XBoost 加速平台,則處理算子優化、量化、編譯和部署等工作。這些資料交代了星凡智能打算怎樣令模型更適合在設備本地推理,卻未提供足以核實 S1 運算吞吐量、功耗或效率提升幅度的數據。246
其他晶片系列按場景劃分:R 系列面向機械人端的多路感測器融合、即時推理和任務決策;K 系列面向衛星在軌推理及遙感數據的本地分析;9 月產品矩陣亦列出供邊緣應用使用的 E 系列,但現有資料不足以判斷它在架構或效能上與其他系列有何分別。245