AI90Training & Inference Acceleration Solution
Hardware-software co-optimization that breaks through the GPU memory wall and unleashes token performance
GPU compute power has grown nearly 60,000-fold exponentially along Moore's Law, while VRAM capacity and bandwidth — the data pathway — have increased only a dozen or so times. This vast gap between the two has created the "GPU memory wall" dilemma.
AI90
Hardware-software co-designed optimization
AI90 is designed to effectively boost GPU utilization by unleashing GPU performance potential and intelligently scheduling the KV Cache. Through lightweight hardware assisted by software optimization, it continuously enhances training and inference potential, empowering users to achieve professional-grade AI capabilities on low-cost hardware.
AI90
Unleashing the Professional Potential of Consumer-Grade GPUs
With intelligent P2P interconnect technology, AI90 boosts GPU parallel computing power by up to 5.6×, while intelligent KV Cache scheduling delivers significant improvements in AI inference capability.
About GenstorAIge
Built on the team's 20+ years of deep IT industry experience and a strong local R&D team, GenstorAIge is dedicated to delivering enterprise-class, industrial-class, and automotive-grade solid-state drives (SSDs), Retimer/Redriver ICs for PCIe Gen5/6, and highly innovative integrated AI training & inference solutions that lead China and meet world-class standards — injecting powerful momentum into the digital transformation of China and the world.
Measured Performance
Time to First Token reduced by 50× and token output bandwidth increased by 5×, with the performance gains growing significantly as model complexity and context length increase.
Through lightweight hardware assisted by software optimization, AI90 continuously enhances training and inference potential, helping users get the most out of expensive compute resources. The acceleration gains grow significantly with model complexity and context length; even lower-end hardware sees substantial speedups, and the acceleration scales linearly with the number of GPUs.
GenstorAIge Technology (Beijing) Co., Ltd. — Building AI-Native storage, to make data more intelligent.