The ThinkSystem AMD Alveo V70 Datacenter Accelerator Adapter is an energy-efficient AI inference card designed specifically for video analytics and natural language processing (NLP) workloads. It supports industry-standard frameworks, allowing for direct compilation of models trained in TensorFlow and PyTorch. The accelerator features AI Engines that create a tightly integrated heterogeneous compute platform, enhancing the performance of NLP and Convolutional and Recurrent Neural Networks (CNN/RNN) for both cloud and edge applications.

The Alveo V70 features the AMD XDNA architecture with 2nd Gen AIE-ML tiles, offering a peak INT8 integer performance of 404 TOPS and a peak BF16 floating point performance of 202 TOPS. It has a memory bandwidth of 47.6 TB/s for internal memory and 76.8 GB/s for external DDR. The accelerator supports up to 96 channels of 1920x1080p video at 10 frames per second using H.264/H.265, operates on a PCIe Gen5 x8 interface, and has a power envelope of 75W.

The Alveo V70 accelerator is well-suited for various AI inference use cases, including video content analysis, natural language processing, and applications in smart city and smart retail environments. Its efficient design and high channel density make it particularly effective for edge applications, helping to reduce costs per AI channel.