๐Ÿ•ท๏ธ
SetupAI
๐ŸŒStalecollected in 58m

Alibaba unveils Zhenwu M890 as NVIDIA alternative

Alibaba unveils Zhenwu M890 as NVIDIA alternative
PostLinkedIn
๐ŸŒRead original on The Next Web (TNW)
#ai-chips#gpu#semiconductors#china-techzhenwu-m890alibabanvidiat-head

๐Ÿ’กA key development in the AI chip war: Alibaba's new GPU aims to bypass US export controls.

โšก 30-Second TL;DR

What Changed

Zhenwu M890 is a new GPU-class AI chip from Alibaba's T-Head.

Why It Matters

This launch highlights China's push for semiconductor self-sufficiency, potentially altering the competitive landscape for AI hardware in the region.

What To Do Next

Evaluate the performance benchmarks of the M890 against NVIDIA A100/H100 if your enterprise operates in the Chinese market.

Who should care:Enterprise & Security Teams

Key Points

  • โ€ขZhenwu M890 is a new GPU-class AI chip from Alibaba's T-Head.
  • โ€ขDesigned as a domestic alternative to NVIDIA chips in China.
  • โ€ขThe chip is already in scaled mass production.

๐Ÿง  Deep Insight

Web-grounded analysis with 26 cited sources.

๐Ÿ”‘ Enhanced Key Takeaways

  • โ€ขThe Zhenwu M890 delivers three times the performance of its predecessor, the Zhenwu 810E, and is specifically engineered for emerging 'AI agents' and agentic AI workloads that demand extensive memory and high-speed communication for complex, multi-step tasks.
  • โ€ขThe chip features 144 GB of GPU memory and 800 GB per second interchip bandwidth, indicating a focus on larger-context, memory-intensive applications like multimodal models and long-context Large Language Models (LLMs).
  • โ€ขAlibaba's T-Head unit has already shipped over 560,000 Zhenwu units (including previous generations) to more than 400 customers across 20 industries, demonstrating significant market adoption and scaled mass production.
  • โ€ขAlibaba has outlined a multi-year chip roadmap, planning to release the Zhenwu V900 in Q3 2027 and the Zhenwu J900 in Q3 2028, with each successor expected to deliver roughly threefold performance gains over its predecessor.
  • โ€ขThe Zhenwu M890 supports multiple data formats from FP32 down to FP4, demonstrating adaptability to the extreme quantization trends prevalent in the large model community.

๐Ÿ› ๏ธ Technical Deep Dive

  • The Zhenwu M890 is a 'training-inference unified AI chip' designed for both AI training and inference tasks.
  • It features 144 GB of GPU memory.
  • The chip provides 800 GB per second interchip bandwidth.
  • It natively supports various data precisions, ranging from FP32 down to FP4.
  • The M890 is paired with Alibaba's proprietary ICN Switch 1.0 interconnect chip, which enables full-bandwidth 128-card connectivity.
  • This interconnect system reduces communication latency to the hundred-nanosecond level, allowing 128 AI chips to operate collaboratively as a single computing unit.
  • The chip is purpose-built to handle the heavy memory and communication demands of agent workloads, where models must retain long stretches of context and coordinate in real time.
  • The predecessor, Zhenwu 810E, was equipped with 96 GB of High Bandwidth Memory 2e (HBM2e) and an inter-chip bandwidth of up to 700 GB per second.

๐Ÿ”ฎ Future ImplicationsAI analysis grounded in cited sources

Alibaba will strengthen its full-stack AI ecosystem.
The launch of the Zhenwu M890, alongside the Qwen3.7-Max LLM and the Panjiu AL128 server system, demonstrates Alibaba's integrated strategy across chips, cloud infrastructure, and AI models.
China's domestic AI chip market will experience intensified competition and accelerated development.
Ongoing US export controls continue to drive Chinese firms like Alibaba and Huawei to invest heavily in self-sufficiency, fostering a robust local ecosystem and increasing competition among domestic players.
Alibaba Cloud's profitability will improve due to the deployment of in-house chips.
Alibaba management has stated that utilizing proprietary chips provides the most cost-effective computing resources, which is expected to enhance the profit margins of its cloud unit.

โณ Timeline

2018-09
Alibaba establishes T-Head (Pingtouge) as its semiconductor subsidiary.
2019-09
Alibaba unveils Hanguang 800, its first AI inference chip, developed by T-Head.
2021-10
Alibaba announces Yitian 710, a 5nm Arm-based CPU for cloud servers, designed by T-Head.
2025-09
China Central Television reports T-Head developed an AI chip (PPU, later identified as Zhenwu 810E) with performance comparable to NVIDIA's H20 and A800 GPUs.
2026-01
Reports emerge about Alibaba preparing T-Head for an independent IPO.
2026-05
Alibaba unveils the Zhenwu M890 AI chip, a successor to Zhenwu 810E, at its annual Cloud Summit.
๐ŸŒRead original article on The Next Web (TNW)
๐Ÿ“ฐ

Weekly AI Recap

Read this week's curated digest of top AI events โ†’

๐Ÿ‘‰Related Updates

Same topic

Explore #ai-chips

Same product

More on zhenwu-m890

Same source

Latest from The Next Web (TNW)

Ctrip fined 5.18B RMB for antitrust violations

Ctrip fined 5.18B RMB for antitrust violations

่™Žๅ—…โ€ขJul 27
The Hidden Costs of Solo Sales Handoffs

The Hidden Costs of Solo Sales Handoffs

The Next Web (TNW)โ€ขJul 28
Hotel groups launch ChatGPT apps for travel booking

Hotel groups launch ChatGPT apps for travel booking

The Next Web (TNW)โ€ขJul 28
Nvidia may guarantee $250bn for OpenAI data center

Nvidia may guarantee $250bn for OpenAI data center

The Next Web (TNW)โ€ขJul 27

AI-curated news aggregator. All content rights belong to original publishers.
Original source: The Next Web (TNW) โ†—