SetupAI
EN
來源較早收集於 26m

平頭哥發布真武M890 AI晶片，性能提升3倍

閱讀原文： 36氪
#ai-chip#hardware#cloud-computing

平頭哥推出全新高性能AI晶片，為大規模訓練集群帶來3倍性能提升。

30 秒速覽

有什麼變化

真武M890內建144GB顯存，片間互聯頻寬達800GB/s。

為什麼重要

M890顯著提升了阿里巴巴在Agentic AI領域的基礎設施能力，為中國市場的高性能計算集群提供了具競爭力的替代方案。

下一步行動

若您在阿里雲生態系內運作，請評估真武M890的規格是否適用於您的大規模模型訓練工作負載。

誰應關注：Enterprise & Security Teams

關鍵要點

  • •真武M890內建144GB顯存，片間互聯頻寬達800GB/s。
  • •性能較真武810E提升3倍。
  • •原生支援FP32至FP4多種數據精度，滿足訓練與推理需求。
  • •配合ICN Switch 1.0晶片，可實現64卡全頻寬互聯。

深度解析

背景與延伸：來自公開資料，非原文內容。引用 18 個來源。

增強重點摘要

  • •The Zhenwu M890 is a core component of Alibaba's 'AI Golden Triangle' strategy, which integrates T-Head's proprietary chips with Alibaba Cloud's infrastructure and the Qwen large language model family to create a comprehensive AI supercomputing system.
  • •The chip is specifically optimized for 'agentic AI workloads,' which are characterized by their demand for extensive working memory to retain context and high-speed communication capabilities for complex reasoning.
  • •Alibaba Cloud has introduced a 128-card supernode server, named Pangu AL128, which incorporates the Zhenwu M890 and the ICN Switch 1.0, achieving communication latency as low as the hundred-nanosecond level for integrated computing.
  • •T-Head has demonstrated significant market penetration, having delivered 560,000 Zhenwu units to over 400 enterprise customers across more than 20 industries as of May 2026.
  • •The Zhenwu M890's predecessor, the Zhenwu 810E, was reported to offer performance comparable to Nvidia's China-compliant H20 accelerator and superior to the Nvidia A800, indicating T-Head's competitive standing in the domestic market.

競品分析

Memory
T-Head Zhenwu M890 (2026)
144GB
T-Head Zhenwu 810E (2026)
96GB HBM2e
Nvidia H20 (China-compliant)
96GB (implied)
Huawei Ascend 910B (2022)
Double on-chip memory of 910, HBM2e (specific capacity not detailed)
Interconnect B/W
T-Head Zhenwu M890 (2026)
800GB/s
T-Head Zhenwu 810E (2026)
700GB/s
Nvidia H20 (China-compliant)
700GB/s (implied)
Huawei Ascend 910B (2022)
Higher bandwidth than 910 (specific value not detailed)
Performance
T-Head Zhenwu M890 (2026)
3x Zhenwu 810E
T-Head Zhenwu 810E (2026)
Comparable to Nvidia H20, surpasses A800
Nvidia H20 (China-compliant)
Comparable to Zhenwu 810E
Huawei Ascend 910B (2022)
Matches/surpasses Nvidia A100 in some tests, 80% A100 efficiency in LLM training, 20% better in others
Data Precision
T-Head Zhenwu M890 (2026)
FP32 to FP4
T-Head Zhenwu 810E (2026)
Not specified
Nvidia H20 (China-compliant)
Not specified
Huawei Ascend 910B (2022)
FP16, INT8 (Ascend 910)
Manufacturing Process
T-Head Zhenwu M890 (2026)
Not specified
T-Head Zhenwu 810E (2026)
Not specified
Nvidia H20 (China-compliant)
Not specified
Huawei Ascend 910B (2022)
SMIC's 7nm fabrication process
Architecture
T-Head Zhenwu M890 (2026)
Proprietary parallel computing architecture
T-Head Zhenwu 810E (2026)
Proprietary parallel computing architecture
Nvidia H20 (China-compliant)
Not specified
Huawei Ascend 910B (2022)
DaVinci cores (25 active cores in 910B vs 32 in 910)
Pricing
T-Head Zhenwu M890 (2026)
Not available
T-Head Zhenwu 810E (2026)
Alibaba Cloud pricing increased 5-34% (March 2026)
Nvidia H20 (China-compliant)
Not available
Huawei Ascend 910B (2022)
Up to 40% lower cost than imported alternatives (for T-Head PPU generally)

技術深入

  • The Zhenwu M890 features 144GB of memory and 800GB/s interconnect bandwidth.
  • It supports diverse data precisions ranging from FP32 down to FP4 for both training and inference tasks.
  • The chip is integrated into Alibaba Cloud's new 128-card supernode server, known as Pangu AL128.
  • The M890 pairs with the ICN Switch 1.0, which facilitates 64-card full-bandwidth cluster interconnects and reduces communication latency to the hundred-nanosecond level.
  • Its predecessor, the Zhenwu 810E, was built on a proprietary parallel computing architecture and utilized in-house inter-chip interconnect technology, complemented by a fully self-developed software stack for end-to-end vertical integration.

前景展望基於引用來源的 AI 分析

Alibaba Cloud's market share in China's AI infrastructure will likely increase significantly.
With its proprietary Zhenwu M890 chip and a full-stack AI Cloud strategy, Alibaba is uniquely positioned to meet the escalating demand for AI compute in China, especially given ongoing global chip shortages and geopolitical pressures.
T-Head is likely to pursue an Initial Public Offering (IPO) in the near future.
Reports indicate that Alibaba is planning an IPO for its T-Head unit to secure substantial capital for further investments in AI infrastructure and to capitalize on the unit's growing market validation and strategic importance.
The development of domestic AI chips like Zhenwu M890 will accelerate China's technological self-reliance in advanced AI capabilities.
Alibaba's substantial investment in T-Head directly addresses the national imperative for domestic alternatives to advanced AI accelerators, driven by U.S. export controls and a strategic push for semiconductor self-sufficiency.

時間線

2018-09
Alibaba establishes T-Head (Pingtouge) as its semiconductor subsidiary.
2019-09
Alibaba announces Hanguang 800, its first self-developed AI inference chip.
2025-09
T-Head develops a new series of AI accelerators (PPU), with performance comparable to Nvidia's H20 and A800 GPUs, deployed by China Unicom.
2026-01-29
T-Head officially unveils the Zhenwu 810E AI chip on its website, confirming performance comparable to Nvidia H20.
2026-03-20
T-Head's Zhenwu 810E GPU reaches large-scale production, with 470,000 units delivered as of February 2026.
2026-05-20
Alibaba's T-Head launches the Zhenwu M890 AI chip, offering 3x performance over its predecessor.
閱讀原文： 36氪

AI 週報

閱讀本週精選 AI 大事摘要 →

PostLinkedIn

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： 36氪 ↗

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。