Alibaba’s semiconductor arm, T-Head, has officially launched its advanced AI chip, the Zhenwu 810E, marking a significant step in China’s pursuit of domestic AI hardware alternatives. The chip boasts performance comparable to Nvidia’s H20, signaling a growing challenge to established global players in the rapidly expanding AI market.
The Zhenwu 810E is a self-developed parallel processing unit (PPU) engineered for both AI training and inference. It features 96 gigabytes of High Bandwidth Memory 2 Enhanced (HBM2e) and offers a substantial inter-chip bandwidth of 700 gigabytes per second. This technical specification positions it as a strong competitor to Nvidia’s offerings, especially the H20, which was specifically designed for the Chinese market under US export restrictions.
Alibaba’s push with the Zhenwu 810E underscores China’s broader strategy to reduce reliance on foreign technology, particularly US-made semiconductors, amidst ongoing geopolitical uncertainties and export controls. The company’s T-Head unit has reportedly surpassed domestic rival Cambricon Technologies in shipments of advanced AI chips, highlighting Alibaba’s growing influence in this critical sector. The chip’s deployment in large clusters on Alibaba Cloud, serving over 400 customers including major state-owned enterprises and tech companies, demonstrates its readiness for real-world, demanding AI workloads.
The launch of the Zhenwu 810E completes what industry experts refer to as Alibaba’s AI “golden triangle.” This strategy integrates three core components: AI model development through Tongyi Lab, cloud services via Alibaba Cloud, and the newly unveiled in-house chips from T-Head. This comprehensive, full-stack approach aims to create a self-reinforcing ecosystem, similar to that of Google, where advancements in one area directly benefit and drive progress in others. This integrated development model allows for optimized performance, cost-effectiveness, and greater control over the entire AI value chain.
While specific performance benchmarks are still emerging, industry insiders suggest that the Zhenwu 810E’s capabilities are comparable to Nvidia’s H20, with potential for future versions to approach or even exceed the performance of Nvidia’s A100. The chip has already been optimized for Alibaba’s own Qwen large language models, indicating its practical application in cutting-edge AI development. The successful deployment and growing customer base suggest that T-Head is well-positioned to capture a significant share of the domestic AI chip market, offering a robust alternative to international suppliers.