PanLink Achieves Multiple Global Firsts in MLPerf® Storage v2.0

UBIX Dominates MLPerf® Storage v2.0, Claiming Seven World Records in AI Storage Performance

In a stunning display of storage prowess, UBIX Technology Co., Ltd. has emerged as a frontrunner in the latest MLPerf® Storage v2.0 benchmark results, released on August 4th at 23:00 Beijing Time. Collaborating with strategic partners Yinfeng Intelligent Computing and Bigbit Micro, UBIX conquered the rigorous testing suite, securing an impressive seven world-leading performance metrics across ten challenging AI workload scenarios.

The MLPerf® Storage v2.0 benchmark, orchestrated by the globally recognized MLCommons® consortium, is the gold standard for assessing storage system performance under demanding AI workloads. Designed to accurately replicate the real-world storage demands of large-scale AI model training, this iteration of the benchmark attracted participation from 26 of the world’s foremost technology innovators.

The diverse field of contestants spanned the storage system, AI infrastructure, and semiconductor sectors, encompassing both innovative players like UBIX and established industry giants such as IBM, Huawei, HPE, DDN, Intel, Samsung, Micron, Kioxia, Lightbits Labs, H3C, and Oracle. This broad participation underscores the MLPerf® Storage v2.0’s widespread recognition and influential role within the global tech landscape.

Seven World Records: A Testament to Unrivaled Performance

Leveraging the AI compute power and network resources of strategic partner Yinfeng Intelligent Computing in Xinjiang, coupled with cutting-edge NVMe SSD technology from Bigbit Micro, UBIX successfully completed all test scenarios. This achievement highlights the exceptional capabilities of the UBIX UbiPower 18000 distributed all-flash storage product line when applied in AI training environments.

The MLPerf® Storage v2.0 benchmark comprises ten distinct testing scenarios, divided into six training tasks and four checkpoint tasks. UBIX utilized a storage cluster consisting of three UbiPower 18000 storage nodes to compete in all ten scenarios, ultimately achieving top honors in the following seven areas:

Training Tasks:

• ResNet50: Achieved a bandwidth of 280.77 GiB/s while supporting 3120 simulated A100 accelerators, securing the top ranking.
• ResNet50: Achieved a bandwidth of 374.57 GiB/s while supporting 2160 simulated H100 accelerators, again claiming first place.
• CosmoFlow: Secured the leading position with a bandwidth of 226.44 GiB/s and support for 608 simulated A100 accelerators.
• CosmoFlow: Topped the charts once more with a bandwidth of 273.21 GiB/s and support for 528 simulated H100 accelerators.
• U-Net 3D: Registered a bandwidth of 455.05 GiB/s while supporting 336 simulated A100 accelerators, earning the #1 spot.

Checkpoint Tasks:

• Llama3-1t Test Model: Delivered a read bandwidth density of 54.7 GiB/s and a write bandwidth density of 36.3 GiB/s per storage system RU (Rack Unit), achieving the highest combined read/write bandwidth density.
• Llama3-405b Test Model: Demonstrated a read bandwidth density of 50.5 GiB/s and a write bandwidth density of 36 GiB/s per storage system RU, once again claiming the leading position in combined read/write bandwidth density.

About UBIX Technology

UBIX Technology Co., Ltd., headquartered in Shenzhen, China, is a technology company specializing in innovative storage products and solutions. Through proprietary innovations in media application, system architecture, and software implementation, UBIX has developed the UBIXFS high-performance distributed file system. Its product portfolio includes the UbiPower series of distributed all-flash storage, the UbiScale series of distributed mass storage, the UbiCube AI-integrated machine, the UbiMatrix high-performance network platform, and the UbiDirector intelligent computing management platform, catering to the full spectrum of high-performance computing (HPC) and AI computing scenarios.

The UBIX UbiPower 18000 distributed all-flash storage solution leverages a unique all-solid-state two-layer resource pool architecture. Powered by RDMA high-speed transmission technology and centered around storage virtualization, it constructs a tiered resource pool of “storage-class memory + flash memory.” This architecture drives overall system performance to unprecedented levels through storage-class memory acceleration, establishing a new benchmark for performance.

UBIX’s AI storage products are commercially deployed in supercomputing and AI computing centers, delivering exceptional performance in data pre-processing, massive data access, and large-scale checkpoint read/write scenarios. They provide efficient storage services for research, industrial HPC, and the development and training of text, video, and multimodal large models.

Looking ahead, UBIX remains committed to the AI storage sector, continuously increasing R&D investment around high performance, high reliability, and intelligence. The company will continue to optimize storage architecture and software capabilities, introducing more leading products and solutions for large-model training and supercomputing and AI computing centers, assisting global users in effectively addressing the storage challenges of the AI era.