InfiniBand, the rising star in the communications industry, has become the focus of the current industry following the development of AI. InfiniBand is a high-performance, low-latency, high-bandwidth communication architecture and network protocol. It was designed in a different way. In order to pursue extremely low latency and high forwarding performance, it bypasses the CPU unit and directly forwards traffic through remote direct memory access technology (RDMA), which greatly reduces the CPU load and improves traffic forwarding performance. At the same time, the traffic does not pass through the CPU and does not need to be repeatedly decoded and encoded, which can reduce the latency to a very low level.
Everything has a price. The new protocol brings high hardware costs and deployment learning costs, which is unacceptable for traditional Ethernet data centers. After all, it is not unusable. However, with the advent of the AI era, the development of artificial intelligence is in full swing, and higher requirements are placed on the latency and performance of computing networks. Traditional Ethernet protocols are no longer applicable. Where there is demand, there is development.
Although Ethernet has also launched RoCE (lossless Ethernet) technology based on RDMA, it still needs to be supported by Ethernet, and Ethernet device hardware also needs to support RoCE. Therefore, InfiniBand is more popular in high-performance computing network data centers. The AI era has just begun, and it is fast step by step. In order to seize the market, the initial additional investment is acceptable.
Infiniband performs RMDA through Infiniband adapters or switches instead of Ethernet. The network must be built through IB NIC and IB switch. Accordingly, it has the following advantages:
InfiniBand supports very high data transmission rates, from the initial 10Gb/s (SDR) to today's 800Gb/s, or even higher, which can meet the large-scale data transmission needs in supercomputers and data centers.
The latency is usually less than 100 ns, while traditional Ethernet is 230ns in specific environments, which is more suitable for AI computing than Ethernet.
Supports automatic path redundancy, error detection and correction mechanisms to ensure the integrity and reliability of data transmission.
Provides RDMA (Remote Direct Memory Access) function, which enables data to be transferred directly between host memory and network devices, bypassing the operating system, thereby significantly improving efficiency.
Today, high-performance computing data centers require networks with large bandwidth and extremely low latency. In this case, traditional Ethernet TCP/IP network protocols cannot meet the needs of data centers because they increase CPU processing resources and high latency. RoCE and Infiniband with RMDA support would be ideal choices, and enterprises would consider their unique requirements and costs when choosing between them. If an enterprise needs the highest performance network connection, Infiniband would be better. For those enterprises seeking optimal performance, easy management and limited costs, RoCE would be more suitable for them.