Weka and Firmus Technologies demonstrated a proof of concept that increases AI token output by 6.5 times by optimizing memory usage and reducing redundant GPU data reprocessing.

Key Points

• Weka and Firmus Technologies achieved a 550% increase in token output using existing GPU infrastructure and power budgets.
• The collaboration utilized Weka’s Augmented Memory Grid on NeuralMesh to extend memory and preserve context for AI agents.
• The technology eliminates the "recompute tax" caused by GPUs repeatedly processing data when memory windows are limited.
• This approach allows organizations to extend the lifespan of existing hardware investments by engineering out system obsolescence.
• The results were presented at the Nvidia GTC AI Conference & Expo to highlight solutions for current data center memory bottlenecks.

Why it Matters

Optimizing AI memory allows companies to scale agentic workloads significantly without the massive capital expenditure required for new hardware or additional data centers. By maximizing the efficiency of existing infrastructure, businesses can accelerate complex tasks like drug discovery and financial trading while maintaining current energy consumption levels.