BerthaE is an AI chip designed to enable Generative and Agentic AI at the edge. Since LLM inference is bottlenecked by memory, BerthaE streamlines the memory access to achieve 90% bandwidth utilization compared to GPU's 50%. It also features optimized dataflow, compute resources, and chip-to-chip networking for efficient LLM inference. BerthaE delivers 3.2x higher performance, 3x better power efficiency, and 10x better cost efficiency than GPUs.

HyperAccel is a fabless company developing AI chips to enable widespread deployment of Generative AI. Therefore, we target inference over training and focus on LLMs instead of traditional AI models. Due to distinct sequential characteristic of LLM inference, GPU designed for other compute-intensive AI workloads is over-specified and thus costly. We overcome the cost barrier by designing the most efficient hardware solution catered to GenAI.