Meet this member at:
China International Semiconductor Executive Summit
NovuMind helps companies put the power of AI in their products and services. Our NovuTensor chip is an industry first, performing tensor computation at the speed of silicon and providing breakthrough performance-to-power ratios. Our full stack of deep learning tools makes it easy to use NovuTensor in a range of applications, from embedded to cloud.
AI will transform many industries, from healthcare to transportation. But there have been barriers to widespread adoption. One has been the computational power demanded by deep learning deployment. Once a neural network model has been trained, it is deployed in an end-user application where it provides “inference” intelligence, such as real-time object detection in video. Running the deployed model in real-time is computationally intensive. This becomes a stumbling block if the app must run in a device with a limited power budget.
An alternative may be to rely on computational acceleration in the cloud, but scaling that to support large numbers of users may also be a challenge. Power and cooling requirements can make it prohibitive to operate inference acceleration at large scale in the data center, with traditional CPU/GPU-based approaches.
Another barrier has been the lack of AI expertise. Few companies have the luxury of an in-house AI team. Model selection, training and tuning are part art and part science, relying on the experience of engineers who have deployed deep learning in real-world applications.
NovuMind’s mission is to eliminate these barriers, so that companies of all sizes, in all industry sectors, can unleash the full power of AI. The NovuMind team has deep expertise in AI, chip design, and high-performance computing. We have 25 team members in USA and 15 in China. Our entire team is passionate about enabling
AI in products and services that will make the world a better place.
深度学习需要较强的计算能力。训练后的神经网络模型会被部署在提供“推理”智能的终端应用中，例如视频中的实时物体检测。实时运行已部署的模型需要密集计算，而终端设备的功耗限制会成为算力的阻碍。另一种选择可能是采用传统的基于CPU / GPU通过云端计算加速，但耗电量和散热要求可能会使其无法在数据中心内大规模运行推理加速。