allworld.mediaNews from all over the world
Alibaba Cloud introduces Apsara supernode stacks integrating Zhenwu V900 accelerators with T-Head components
Read more
Pandaily
pandaily.com

Alibaba Cloud introduces Apsara supernode stacks integrating Zhenwu V900 accelerators with T-Head components

At the Apsara Cloud and T-Head conference, an updated supernode stack was demonstrated. This stack integrates the Zhenwu V900 artificial intelligence accelerator with ICN Switch chips from T-Head, the Pangu NIC intelligent network card, and the Zhenyue SSD controller.

The main goal of the presentation was not to showcase an individual accelerator, but to demonstrate the coordinated operation of computing, networking, and system components at the entire supernode level. The official product branding is defined as Alibaba Cloud / T-Head.

According to QuDong's Apsara report, the Zhenwu V900 accelerator features 216 GB of memory and provides an inter-chip connection speed of 1200 GB/s. Its performance is approximately three times higher than that of the Zhenwu M890, and it supports native FP8/FP4 formats. Mass production of the Zhenwu V900 is planned to begin in the first quarter of 2027.

Thanks to the connection via ICN Switch, the supernodes achieve unified addressing and native memory semantics, allowing approximately one thousand V900 cards to function as a single large accelerator. Using Alibaba Cloud's repurposed intelligent computing network, the company states that one cluster can scale up to about 500,000 cards.

Previously, M890-based supernodes were used on a large scale and, according to the company, processed models such as Qwen3.8 and Kimi K3, which have over two trillion parameters, although this has not been independently verified.

This development is significant because T-Head now positions AI accelerators, Yitian processors, ICN interconnects, Pangu NIC network cards, and Zhenyue storage controllers as a unified set of chips for domestic data centers.

Although subsequent generations of Zhenwu and supernode SKUs are planned for 2029, the information from September 22 specifically concerns the V900-based supernode and the potential of a 500,000-card cluster.

For cloud service and cluster buyers, this news means that Alibaba is packaging the accelerator, switch, network card, and silicon SSD components into a single Apsara supernode concept with a clear scalability metric.

The delivery speed of the V900 (stated as Q1 2027) and the actual behavior of the multi-stack factory connection will determine whether the 500,000-card path is an engineering reserve or a near-term plan.

Similar stories

T-Head Unveils Yitian Server Processor Roadmap for 2027, Including ICN-Link Support
Read more
pandaily.com

T-Head Unveils Yitian Server Processor Roadmap for 2027, Including ICN-Link Support

Alibaba's chip division, T-Head, presented its roadmap for the Yitian processor line during the Apsara Conference in Hangzhou in 2026. According to the information presented, the Yitian 720 and Yitian 730 models are scheduled for release in 2027. Furthermore, the later Yitian 750 model is expected to feature T-Head's ICN connection, enabling direct connection of Yitian processors to Zhenwu AI accelerators.

According to coverage of the event by QuDong and IT Home via Phoenix Tech, the Yitian 720 processor will focus on improving single-core performance, core density, and energy efficiency. The Yitian 730 model will be the first in the Yitian series built on T-Head's fully self-developed processor microarchitecture, with a claimed single-core SPECint2017/GHz performance that could reach approximately 1.4 times that of the Yitian 710.

In an additional report from Chipsee/ifeng covering the same key speech, sketches of SKU-level roadmap were presented: the Yitian 720 is anticipated with around 192 cores, 12-channel DDR memory, and 96-lane PCIe, demonstrating a SPECrate performance approximately 1.1 times higher than the 710 at the same frequency. The Yitian 730, in turn, is planned with about 32 self-developed cores, SMT2 support, 8-channel DDR, and 128-lane PCIe, with performance ranging from 1.25 to 1.4 times higher.

These figures should be viewed as plans rather than finalized technical specifications, as they were presented on roadmap slides. Regarding timelines, the secondary report indicates that both models (720 and 730) are slated for release in the third quarter of 2027.

The longer horizon covers the Yitian 750, described as the second self-developed processor supporting ICN-link. This feature will allow direct connection between CPU chips and Zhenwu AI chips, bypassing traditional PCIe connections. T-Head's broader step allows linking high single-threaded performance and memory bandwidth processors with agent workloads that still rely on host processors for scheduling, state management, and data preparation near the accelerators.

It is important to note that parameters such as core count, DDR channels, and SPECrate multipliers are derived from secondary coverage of the Apsara Conference and should be considered targets until T-Head publishes final technical datasheets. Only the public direction of the Yitian 720/730 roadmap for 2027 and the concept of Yitian 750 with ICN connection to Zhenwu have been confirmed for this review, not specific release dates or benchmarking results.

Popular