ShenzhenJune 22, 2026 /PRNewswire/ — Recently, Union Memory, in collaboration with Intel and H3C, jointly released a new generation of distributed storage solutions tailored for AI computing scenarios. Based on H3C’s AI-native storage UniStor X20000, the solution is equipped with Intel® Xeon® 6 processors and Union Memory’s UH812a enterprise-grade PCIe 5.0 SSD. Through multi-level technological innovation in storage-computing synergy, it systematically addresses the complex challenges of high throughput, high bandwidth, low latency, and coexistence of multiple data types in AI training and inference, building a robust data storage foundation for large models with hundreds of billions of parameters.
This storage solution, powered by Intel® Xeon® 6 processors, meets the “three highs” standards of high performance, high efficiency, and high reliability, confidently tackling the challenges of massive data.
- Architectural Performance Breakthrough
Utilizing RDMA and SPDK full user-mode technology stack, achieving zero-copy data interaction, combined with distributed global cache, unleashing extreme performance at the architectural level. - Unified Storage for All Scenarios
Enables lossless interoperability of unstructured protocols, providing a scalable unified storage solution. A single architecture simultaneously meets the storage needs of multiple data types, including AI computing, structured data, and massive unstructured data. - Elastic Scaling, Business Agility
Streamlined architecture supports linear growth of performance and capacity with cluster size, flexibly matching the elastic demands of business development; end-to-end data protection ensures business continuity. - Intelligent Operations, Secure and Trustworthy
Supports rapid batch deployment of clusters and flexible data distribution strategies, allowing users to customize fault domain isolation and storage location selection, ensuring data reliability and access security.
Hardware-Software Synergy, Building a High-Performance Storage Foundation
This solution is based on a 3-node H3C AI-native storage UniStor X20000, with each node equipped with 2 Intel® Xeon® 6 GNR-SP 6530P processors, 18 Union Memory UH812a PCIe 5.0 NVMe SSDs, and interconnected via 400GbE CX-7 RDMA network cards. The storage system employs RDMA and SPDK full user-mode technology, achieving zero-copy data interaction between modules, combined with a distributed global cache mechanism to maximize hardware potential.
Measured data proves the solution’s solid performance. Under a 3-node storage cluster with 2.4TbE network configuration:
- Impressive Large Block Data Throughput — 1M sequential read throughput reaches 260GB/s, and with EC 4:2 ratio, 1M sequential mixed read/write achieves an extreme throughput of 317GB/s, clearing the path for AI large model data loading;
- Ultra-Fast Small IO Response — 4K random read end-to-end average latency is strictly controlled within 500µs, with cluster IOPS approaching 6 million; 4K sequential write average latency is below 2ms, with IOPS near 1.5 million.
Whether handling large block data transfers or intensive small IO access, this solution delivers outstanding results with industry-leading data performance.
Tripartite Collaboration, Directly Addressing Computing Bottlenecks
As AI large model parameter scales continue to expand, traditional storage-computing integrated architectures can no longer meet the stringent demands of AI computing scenarios. The training phase requires frequent reading of massive small files and writing of large checkpoints, putting dual pressure on storage bandwidth and IOPS; in the inference phase, the access latency of critical data like KV Cache directly impacts GPU utilization, with expensive computing resources often idling due to storage waits. Meanwhile, the coexistence of multiple data types such as block, file, object, and HDFS exacerbates protocol silo issues and increases management complexity. The storage system urgently needs a comprehensive overhaul from architecture to protocols.
Union Memory, together with Intel and H3C, has delivered an “end-to-end collaborative” solution — with Union Memory’s UH812a SSD as the performance cornerstone, Intel® Xeon® 6 processors as the computing hub, and H3C’s AI-native storage UniStor X20000 as the system carrier, achieving deep synergy from media to processors to the storage software stack:
- On the storage media side, Union Memory’s UH812a PCIe 5.0 SSD, with its self-developed controller architecture and extreme reliability, provides a solid foundation for high-speed data flow, fully unleashing the bandwidth potential of PCIe 5.0.
- On the processor side, Intel® Xeon® 6 processors feature built-in AI acceleration engines, with memory bandwidth increased by up to 2.3 times and comprehensive I/O capability upgrades, efficiently serving as the computing hub.
- On the storage system side, H3C’s AI-native storage UniStor X20000, leveraging its self-developed unified storage engine, integrates block, file, object, and HDFS protocols into one, enabling lossless interoperability between unstructured protocols and flexibly handling mixed workloads with multiple data types in AI training and inference.
Additionally, the solution supports linear scaling of up to thousands of nodes, with capacity reaching EB levels and performance growing synchronously with cluster size; end-to-end data protection and flexible fault domain isolation mechanisms ensure business continuity and data security. Compared to configurations based on the previous generation Xeon® EMR 6530 processors (with equivalent SSDs, network cards, and client environments), the new generation solution achieves up to a 25% performance peak improvement in key scenarios.
Based on 3-node solution configuration: GNR-SP 6530P VS EMR 6530
UH812a Hardcore Empowerment, Building a Performance Foundation
As the core performance cornerstone of this solution, Union Memory’s UH812a enterprise-grade PCIe 5.0 SSD uses a self-developed 7nm controller chip, supports the NVMe 2.0 protocol and dual-port design, with sequential read/write speeds up to 14900/10500 MB/s and 4K random read/write performance reaching 3500K/520K IOPS, providing robust support for high-speed AI data flow.
On the reliability front, the UH812a is equipped with an LDPC+DSP error correction algorithm engine, extending flash memory lifespan by 5 times; dynamic power regulation technology keeps sequential read peak power consumption within 18W and standby power consumption as low as 5W. The product has passed Intel’s comprehensive verification and neutron irradiation testing, with an MTBF exceeding 2.5 million hours and an annual failure rate below 0.35%, maintaining stable output even in demanding environments like high-load, long-duration AI cluster operations.
Currently, the Union Memory UH812a has been fully validated in mainstream AI scenarios such as AI training, multimodal data processing, real-time inference, large model loading, and edge computing, accumulating extensive deployment experience and fully matching the diverse business needs of the AI-native storage UniStor X20000.
Union Memory and Intel: Joint Solution Creating Value
As Intel’s first domestic SSD strategic partner in China, Union Memory’s UH812a product has been deeply adapted to the Xeon platform. The two parties have previously jointly launched high-performance storage solutions based on Xeon® 6 and UH812a, as well as RDMA+NVMe network storage solutions, fully validating the significant effectiveness of hardware-software synergy in high-throughput, low-latency scenarios.
In the H3C AI-native storage UniStor X20000 solution, the collaborative innovation between Union Memory and Intel has been validated at the system level. The solution helps users quickly replicate and deploy in scenarios such as AI computing, big data analytics, data warehouses, and enterprise core applications.
This time, Union Memory, together with Intel and H3C, officially released a joint storage solution for AI computing scenarios. Measured data strongly confirms the technical path of “strengthening computing with storage” — that is, using high-performance storage to fully unleash expensive GPU computing power, effectively eliminating data read/write bottlenecks, and ensuring that every bit of computing investment delivers value. Looking ahead, Union Memory will continue to deepen storage technology, enhance collaborative innovation with ecosystem partners, and empower the accelerated deployment of AI across various industries with a solid storage foundation.