Recap of Huawei Connect 2025

Huawei ran last week its 10th Connect conference and expo in Shanghai, China, and it was, once again, a huge event with approximately 15,000 attendees coming from all over the world. This flagship event is the focal point for the Chinese firm to unveil, each year, its current products development and strategic directions. And the audience was not disappointed with the tons of new things, in the infrastructure, data storage, networking and computing domains.

And for several years now, AI is at the center of the expo floor and on stage with new products iterations. The company spend time to illustrate the evolution of data processing and the need for faster results coupled with bigger storage and high bandwidth networks.

The first keynote, presented by Eric Xu, rotating and acting chair of Huawei, was impressive. He unveiled the processors and accelerators strategy with the Ascend and Kunpeng product line and the roadmap is ambitious with several key dates in the coming quarters, putting real pressure on competition. We understand that configurations are massive delivering leading results in various dimensions. It is the case for Atlas A3 SuperPoD and SuperClusters. For instance so far, the CloudMatrix384 has been sold 300+ times to 20+ customers.

Click to enlarge

GPU-oriented computing with Huawei Ascend accelerators illustrated above are not orphan as the company also presented Taishan computing racks fueled of Kunpeng CPU processors. And it is even more impressive when you couple both entities and multiply them.

Click to enlarge

Fundamental element of these architectures to deliver super high performance, the interconnect and globally networking are obviously essential elements and the next strategic news was about UnifiedBus 2.0. To illustrate the capability, the goal is to connect 10,000 NPUs.

Click to enlarge

This networking development marked also a clear direction towards open hardware and open source as the full specifications are available for everyone. It is extended by the Xinghe AI Fabric 2.0 also announced during the conference.

The firm also presented entry level processing cards like the Atlas 350 card or the Atlas 850/860 server named the SuperPoD server.

On the data storage side, the product team insisted on the new-gen all flash OceanStor Dorado product, also known as v7. As a real unified storage platform coupling block, file and object protocols and access methods, it introduced any-to-any NVMe architecture with RoCEv2 internal networking coupled with DPU powered by FlashLink 3.0 new model. A disk enclosure could be coupled with 32 controllers vs the competition and previous approaches thanks to SmartMatrix 4.0. Dorado also introduced FlashEver, a way to easily migrate from previous Dorado editions. With the new design, the platform reaches 99.99999% reliability with RPO = RTO = 0.

Click to enlarge

The A Series line dedicate to AI has a new companion, the A600, more accessible and more oriented towards inference.

The engineering group displayed UCM – Unified Cache Manager – a way to integrate and support any flash storage for KVcache and other needs.

Click to enlarge

SSD has been refresh as well as the 245TB SSD was demonstrated on the sow floor for capacity needs fueled of QLC NAND.

We also discovered the OceanStor Pacific 9926, the companion of 9928, with 4PB in 2U and a compression card. OceanProtect was also highly visible representing the company’s approach to data protection both on the software aspect but also on the backup target with X and E product line.

As always it was an impressive conference with several new products and technology directions.