Solidigm opens AI Central Lab to accelerate AI storage research

Solidigm has announced the launch of its AI Central Lab, featuring what the company claims are the industry’s highest-performing and most dense storage test clusters designed for artificial intelligence workloads. Located in Rancho Cordova at the FarmGPU facility near Solidigm headquarters, the lab enables real-world testing on NVIDIA B200 and H200 GPUs, providing insights into how storage can enhance cluster efficiency.

Click to enlarge

High-Performance and Dense Storage Clusters

The AI Central Lab boasts several notable configurations:

• Performance Cluster: Using Solidigm D7-PS1010 SSDs, it achieved 116 GB/s per-node throughput in MLPerf Storage testing—the highest ever recorded in AI model training.
• Density Cluster: Packs 192 Solidigm D5-P5336 SSDs with 122 TB each, delivering 23.6 PB of storage in just 16U of rack space.

Additional hardware includes 800 Gbps Ethernet networking and high-performance storage servers. Available workloads for testing include AI model training and inference, KV cache offload, VectorDB tuning, and power consumption benchmarking.

Real-World Testing and Research

The lab was built with common reference architectures in mind, allowing customers to test their solutions with Solidigm SSDs under the same standards as large-scale data centers worldwide. This facility enables research beyond conventional component-level testing, providing insights at the server, rack, and cluster levels. Work in the lab is already underway with multiple organizations.

Collaborations and Innovations

In collaboration with Metrum AI, Solidigm demonstrated a technique that offloads AI data from memory to SSDs, reducing DRAM usage by up to 57% during inference with RAG (retrieval-augmented generation) while maintaining performance and accuracy. The partnership with FarmGPU also supports ongoing efforts to optimize AI infrastructure for both performance and efficiency.

Source: Solidigm