Weka launches new storage platform to boost AI efficiency

AI storage company Weka has released its NeuralMesh 6 software alongside its first self-designed hardware line, Wekapod 3. The new platform is designed to enhance AI model performance and efficiency by significantly reducing the reliance on expensive Graphics Processing Unit (GPU) memory.

The platform utilizes Weka's Augmented Memory Grid approach, which aggregates flash storage to function like GPU memory at a fraction of the cost. This aims to address performance bottlenecks in AI models caused by long context windows and multi-turn conversations, which lead to redundant computations and consume valuable GPU resources.

NeuralMesh 6 introduces several new capabilities, including composable and virtual multi-tenancy, offering hardware-level isolation and scaling to thousands of tenants per cluster. It also provides unified file and object storage without separate data copies or translation layers, accelerating data handling, particularly in GPU cloud environments.

Metadata-first replication and AlloyFlash technology further improve performance and cost-effectiveness. AlloyFlash mixes different types of NAND flash memory, allowing the use of cost-efficient QLC memory for bulk data while maintaining speed for latency-sensitive tasks using faster TLC memory.

Weka co-founder and CEO Liran Zvibel highlighted that the new platform enables hundreds of times more efficient caching, eliminating the need for repeated calculations and saving significant GPU resources. This can accelerate the deployment of new AI workloads and improve the utilization of existing GPU investments.