At ISC 2026

At ISC 2026, the European conference dedicated to supercomputing, Huawei presented a new integrated solution for computing, storage, and networking for HPC and AI, enabling the implementation of large-scale AI applications and catering to the needs of universities and research institutes.

Huawei Strengthens Its HPC and AI Offerings at ISC 2026

The new offering from Huawei is named AI Data Center and primarily focuses on data management throughout its life cycle. Huawei notes that data management platforms in research (RDM, Research Data Management) are converging towards the widely used data lake model for AI, which is why it proposes AI Data Lake, a platform based on OceanStor Pacific, offering as much as 11 PB of capacity in a 2U form factor. This allows for the storage of a significantly large amount of data in a very compact volume. The solution integrates with DME Omni-Dataverse, enabling centralized and global management of data collection from multiple sites and types within the data lake.

The Huawei OceanDisk 1610 Smart Disk Enclosure is a storage solution designed for the HPC domain, achieving a bandwidth of 220 GB/s and a capacity of 4 PB in a 2U form factor. This solution reduces the occupied space while maintaining high performance for scientific computations and AI training.

Regarding inference, Huawei presents its AI Data Platform “3+1”, which integrates a knowledge base, a KV (key-value) cache, memory, and unified data management functionalities, along with UCM technology to manage in-memory data intended for inference, achieving over 95% accuracy in information retrieval.

Huawei CMS (Context Memory Storage) is a platform specifically designed to accelerate KV caches and support the storage of AI's contextual memory. It integrates with infrastructures based on both GPU and NPU and ensures compatibility with RoCE and Unified Bus network protocols. According to Huawei, compared to traditional solutions, it offers a 3 to 5 times increase in inference throughput, reduces first token latency by up to 90%, and lowers costs by 30%.

Looking at the network, the Xinghe AI Data Center Network is based on Open Ethernet and offers configurations up to 128x 800 GbE or 128x 400 GbE cooled by air or liquid, enabling four times the scale compared to industry standards and reducing TCO by 40% at equal sizes. The optical modules StarryLink introduce load balancing at the single packet level, achieving a network throughput of 98% and a 7% improvement in the efficiency of AI training and inference. The system offers integrated telemetry with sub-millisecond granularity to quickly identify any issues.