來源較早收集於 8m

阿里巴巴推出中國頂級AI模型RISC-V伺服器晶片

阿里巴巴推出中國頂級AI模型RISC-V伺服器晶片
PostLinkedIn
🇬🇧閱讀原文: The Register - AI/ML
#ai-hardware#server-chip#risc-v-chinaalibaba-risc-v-server-chipalibabarisc-v

💡阿里巴巴最速RISC-V晶片跑中國頂AI—挑戰Nvidia/ARM?

⚡ 30 秒速覽

有什麼變化

阿里巴巴推出最強RISC-V伺服器晶片

為什麼重要

推動RISC-V在AI基礎設施應用,提升中國在美限制下的AI硬體自給自足。實現國內LLM如Qwen在開放ISA伺服器高效運行。

下一步行動

針對Qwen2或DeepSeek模型,在阿里巴巴RISC-V晶片上進行推理基準測試。

誰應關注:Enterprise & Security Teams

關鍵要點

  • 阿里巴巴推出最強RISC-V伺服器晶片
  • 專為中國領先AI模型優化
  • 創RISC-V處理器性能新紀錄
  • 落後西方晶片數年

🧠 深度解析

本篇為 AI 生成分析,非原文內容。

🔑 增強重點摘要

  • The chip, identified as the XuanTie C930, utilizes a multi-core architecture specifically designed to overcome memory bandwidth bottlenecks common in large language model (LLM) inference.
  • Alibaba is leveraging its open-source T-Head ecosystem to encourage domestic adoption, aiming to reduce reliance on restricted high-end GPUs from Western suppliers like NVIDIA.
  • The processor incorporates custom vector processing units (VPUs) that provide hardware-level acceleration for specific quantization formats (INT8/FP8) favored by Chinese AI developers.
📊 競品分析▸ Show
FeatureAlibaba XuanTie C930NVIDIA Grace Hopper (GH200)Ampere Altra Max
ArchitectureRISC-VARM NeoverseARM Neoverse
Primary FocusAI Inference (Domestic)AI Training/InferenceCloud Computing
PerformanceHigh (RISC-V record)Industry LeadingHigh (General Purpose)
AvailabilityChina-restrictedGlobalGlobal

🛠️ 技術深入

  • Architecture: 64-bit RISC-V high-performance core design.
  • Vector Extensions: Custom implementation of RISC-V Vector (RVV) 1.0 extensions for AI acceleration.
  • Memory Interface: Support for HBM3 to address high-bandwidth requirements of LLMs.
  • Interconnect: Proprietary high-speed chip-to-chip interconnect for multi-socket scaling.
  • Manufacturing: Fabricated on advanced domestic node (likely 7nm or 5nm class) to bypass export controls.

🔮 前景展望基於引用來源的 AI 分析

Alibaba will increase its market share in the Chinese domestic server market by 15% by 2027.
The shift toward sovereign silicon allows Alibaba to offer a stable, non-embargoed alternative to Western-designed processors for state-backed AI projects.
RISC-V will achieve parity with ARM in server-side AI inference performance within 36 months.
The rapid iteration cycle of the XuanTie series demonstrates that RISC-V is closing the architectural gap faster than historical industry projections.

時間線

2019-07
Alibaba T-Head releases the XuanTie 910, its first high-performance RISC-V processor core.
2021-10
Alibaba open-sources the XuanTie RISC-V IP cores to accelerate the domestic ecosystem.
2023-08
Alibaba introduces the C920, marking a significant step toward server-grade RISC-V capabilities.
2026-03
Alibaba unveils the C930, the most powerful RISC-V server chip to date.
📰

AI 週報

閱讀本週精選 AI 大事摘要 →

👉相關動態

AI 策展新聞聚合。所有內容版權歸原始發布者所有。
原始來源: The Register - AI/ML

這是摘要,不是原文。去看原站,或訂閱每週簡報。

每週電子報

每週一封,可隨時退訂。