The Jotunn8 AI Accelerator by VSORA is a game-changing product in the realm of AI inference, designed to handle any algorithm on any host processor, offering unparalleled programmability. This AI accelerator provides a substantial 6,400 Tflops of performance using fp8 Tensor Cores and is highly adaptable for large language models like GPT-4, significantly reducing deployment costs to below $0.002 per query. Its architecture allows large-scale AI models to function efficiently, emphasizing low latency and minimized power consumption.
Utilizing high-level programming, the Jotunn8 is algorithm-agnostic, meaning it can seamlessly process both AI and general-purpose tasks, chosen layer-by-layer. It is equipped with 192 GB of on-chip memory to support hefty data handling requirements, ensuring that substantial AI workloads can be managed effectively without reliance on external memory systems. This characteristic is crucial in overcoming the 'Memory Wall' challenge inherent in traditional computing setups.
Designed for both cloud and on-premise applications, the Jotunn8’s peak power consumption is pegged at 180W, reinforcing its position as a high-performance yet energy-efficient solution. This AI accelerator provides a balance between energy efficiency and performance, making it an exemplary choice for environments demanding rapid AI deployment and execution.
