Leveraging AI workloads or applications has taken place for many years; however, only recently has it captured the public's attention. Similar to other technology, AI is only useful when implemented with the right solutions. In this case, for AI, the network infrastructure (or fabric) is key. As such, its implementation deserves critical attention.
The following traffic patterns characterize GenAI workloads:
- Extreme data exchange volume
- High-rate data exchange
- Latency sensitivity
- Diverse traffic patterns (ordered and predictable, with peaks and valleys)
- Heterogenous traffic size flows (elephant and mice flows)
The following are strict infrastructure requirements that separate GenAI workloads from the rest such as virtualization, or storage. GenAI workloads must:
- Run on a lossless fabric
- Have access to a high-performance fabric
- Scale according to its needs
There are other types of workloads that impose similar network requirements, but they do not necessarily apply to all three characteristics shown above. For example, while multicast or virtualization traffic workloads require high performance to accommodate the large and bursty amount of traffic generated, it also requires scalability to maintain and service the multicast or virtual environment. However, it can be lossy as the upper layer of the application can perform traffic recovery if needed.