Solidigm Unveils AI Central Lab, Home to Performing and Dense Storage Test Clusters

Solidigm announced the opening of its AI Central Lab featuring an industry’s highest-performing and most dense storage test clusters ever built for AI workloads.Hosted in Rancho Cordova, CA, at the FarmGPU facility near Solidigm headquarters, the AI Central Lab enables the company to run and study real-world workloads on leading-edge Nvidia B200 and H200 GPUs. This provides state-of-the-art insights into how leading storage can optimize cluster efficiency and unlock the ability for key Solidigm AI ecosystem collaborators to run experiments on the lab’s world-class hardware.

“Our Solidigm AI Central Lab combines today’s most powerful GPUs with leading storage infrastructure to unlock new levels of testing and joint innovation for our customers and the developer community,” said Avi Shetty, senior director, ecosystems and partnerships, Solidigm. “These were capabilities previously only available to select companies, and Solidigm is now enabling them while demonstrating the criticality of having storage close to the GPU.”

By building the lab with common reference architectures in mind, the company will enable customers to test their solutions with the firm’s SSDs in an environment using the same standards as large-scale data centers around the world. The lab will also accelerate storage research beyond conventional component-level testing to better comprehend real-world usage scenarios at the server, rack, and cluster levels.

“Running storage tests isn’t enough anymore. In our AI Central Lab, we can execute real-world AI workloads and use our cutting-edge telemetry capabilities to optimize systems for performance and efficiency and gain insights into the storage needs of emerging workloads,” Shetty added.

AI Central Lab features the latest in high-performance compute hardware. Highlights include

• Highest-performing storage test cluster: The Solidigm D7-PS1010 SSD, the performance cluster achieved the highest-ever per-node throughput measured in the MLPerf Storage (AI model training) test at 116GB/s. This can be flexibly scaled to multiple nodes.
• Most dense storage test cluster: Outfitted with 192 D5-P5336 SSDs with a massive 122TBs of storage each, this cluster packs 23.6PBs into just 16U of rack space.
• Additional technology: The lab features Nvidia B200 and H200 GPUs, 800Gb Ethernet networking, and storage servers from leading vendors. Solidigm is also constantly expanding and exploring new technologies.

Workloads available for testing at the lab include

• AI-specific and emerging workloads such as actual, GPU-intensive AI model training and inference and innovative approaches like KV cache offload and VectorDB tuning.
• The lab can benchmark power consumption of different configurations, developing architectures optimized for the most precious resource in AI today.
• The lab can study how to quickly feed data to expensive GPU accelerators, keeping them as busy as possible.
• The AI Central Lab’s data can help translate SSD specifications into system-level, industry-relevant AI efficiency metrics such as tokens per dollar and tokens per watt.

Work in the lab is under way with multiple organizations. An example of the AI Central Lab in collaboration includes Solidigm and experts at Metrum AI delivering a technique that offloads AI data from memory to SSDs to reduce DRAM usage by up to 57 percent during inference with RAG (retrieval-augmented generation).

The collaboration with FarmGPU also allows Solidigm to support their ongoing efforts to optimize AI infrastructure for performance and efficiency.

Resource:
Solidigm AI Central Lab