AMD Pensando™ Pollara 400

The Pollara 400 is a high-performance network card designed to improve data transfer between GPUs in back-end AI networks. It supports 400 Gbps RDMA Ethernet and helps AI systems work faster and more efficiently by meeting the high demands of bandwidth, low latency, and reliable communication.

Traditional Ethernet struggles with AI's specific needs, such as handling large amounts of data with low delays and avoiding congestion. The Pollara 400 solves these problems, ensuring fast and smooth communication across multiple GPUs with features like congestion control and reduced latency. It also offers flexibility, allowing users to work with various vendors and customize their network setups.

Key Features:

P4 Programmability:
The Pollara 400 uses P4, a flexible programming architecture, allowing customization and future updates to match new AI and networking standards. This ensures the card stays relevant and adaptable to changing needs.

Multipathing & Smart Packet Handling:
It efficiently splits and routes data across multiple paths to prevent network congestion and make full use of bandwidth. This results in faster data delivery and lower delays, even in complex AI models. It works seamlessly with AMD Instinct™ GPUs and EPYC™ CPUs while simplifying network setups.

In-Order Data Delivery:
The Pollara 400 handles packets that arrive out of order, ensuring they are correctly processed and sent to the GPU without delays. This reduces latency and keeps the system running smoothly.

Selective Retransmission for Fast Recovery:
Instead of resending all data when a packet is lost, the Pollara 400 resends only the missing packets. This saves bandwidth, reduces delays, and ensures efficient data flow, making it ideal for large AI workloads.