提交成功
BACK

32TB High-Performance SSD Continues to Accelerate AI Large Model Training!

2025 / 03 / 27
0

The rapid advancement of AI large models is driving an unprecedented surge in data growth, pushing data center storage capacity and I/O performance to new limits. While QLC SSDs offer higher storage density, they still fall short of TLC SSDs in performance. Is there a solution that combines both high capacity and high performance? The PBlaze7 7940 30.72TB SSD rises to the challenge, delivering the best of both worlds.

 

 

The PBlaze7 7940 Series PCIe 5.0 SSD is designed for demanding enterprise workloads, delivering powerful support for AI training, big data, cloud computing, virtualization, and more. The 30.72TB model, built with TLC NAND technology, offers up to 14GB/s sequential read and 9.7GB/s sequential write bandwidth, along with 2.7 million IOPS for 4K random reads and 410K IOPS for steady-state 4K random writes—maintaining a leading position among same-capacity products!

 

Compared to high-capacity QLC SSDs, the PBlaze7 7940 30.72TB offers superior read performance and up to 3X higher write performance, making it an optimal choice for AI large model training tasks such as sample data storage, data cleansing, and checkpoint saving. It reduces write operation time by nearly 70%, significantly accelerating GPU recovery for training resumption and optimizing the overall training process.

 

In the MLPerf Storage Benchmark for machine learning storage performance, a single PBlaze7 7940 30.72TB SSD, tested with the UNet-3D model and eight simulated A100 GPUs, achieved a training speed of 9,727MB/s—76.3% higher than mainstream PCIe 4.0 SSDs and surpassing a competitor’s 7.68TB PCIe 5.0 SSD, which reached 9,258MB/s.

 

 

 

 

The transition from PCIe 3.0 to 4.0 took nearly six years, while the shift to PCIe 5.0 happened in less than three! During this time, AI has driven an explosive increase in computing power, with single-GPU training performance soaring from tens to thousands of TFLOPS. Advancements such as compute-storage disaggregation, high-performance networking, and specialized data processing units have continuously eliminated bottlenecks between storage and computation. In this landscape, maximizing single-drive performance is key to fully unleashing computing potential, while higher capacities are essential for meeting the massive storage demands of ever-expanding model parameters.

 

For instance, a typical 2U server with a 12-drive configuration upgraded to PBlaze7 7940 30.72TB SSDs can provide 368TB of storage per server, with read and write bandwidths exceeding 160GB/s and 118GB/s, respectively. This increase in both storage capacity and performance significantly reduces the number of storage nodes needed, saving valuable rack space and allowing clients to deploy more compute units while maintaining high utilization rates.

 

 

 

 

The PBlaze7 7940 30.72TB uses a 2.5-inch U.2 form factor and supports NVMe 2.0, NVMe-MI 1.2b, and TCG Opal specifications. It offers enterprise features such as full data path protection, power loss data protection, AES-256 self-encryption, secure erase, firmware upgrade without reset. Designed for high-speed data caching, edge computing, and distributed storage scenarios, the PBlaze7 7940 series also offers HHHL AIC, EDSFF E1.S, and E3.S form factors.

 

Currently, the PBlaze7 7940 30.72TB SSD is being used by leading AI customers. Memblaze will continue to push the limits of technology, delivering larger capacities, higher performance, and more efficient storage solutions to the market.