Home / Technology / Weka Unlocks More GPU Power with Cheaper Storage
Weka Unlocks More GPU Power with Cheaper Storage
22 Jul
Summary
- Weka extends GPU memory using affordable flash storage.
- New software and hardware integrate to boost AI performance.
- This solution aims to lower inference costs and speed deployments.
GPU memory, a critical and costly resource in AI production, is now being augmented by Weka's new NeuralMesh 6 software and Wekapod 3 hardware. Weka's Augmented Memory Grid approach leverages affordable NAND flash storage to emulate GPU memory at a fraction of the price. This advancement aims to enhance the utilization of existing GPU investments, reduce inference costs, and accelerate the deployment of new AI workloads, addressing the industry's pressing need for more accessible compute power.
The NeuralMesh 6 introduces several key features, including composable and virtual multi-tenancy for improved isolation and rapid provisioning of up to 1,000 tenants per cluster. It also offers unified file and object storage, allowing direct data access through both paths without duplication. Furthermore, metadata-first replication and AlloyFlash technology, which intelligently combines TLC and QLC NAND flash, optimize performance and reduce cost per terabyte. Weka's solution directly tackles the problem of wasted GPU compute from repeated calculations in long context windows and multi-turn conversations by caching pre-calculated tokens.
