Between the vast and random traffic pattern exhibited by GenAI workloads, it is imperative for the network to be lossless. The number of data packets lost while undergoing training, fine-tuning, and inferencing must be lossless. Any amount of data lost can have an adverse effect on the results, whether that be a text, image, video, speech, data augmentation, or others.
To implement a lossless fabric, the fabric supporting a GenAI workload needs to have the right feature set, either through quality of service, enhanced queueing mechanism, or new technology.