Many computer technologies have come and gone, but Ethernet has been around for half a century. Or more precisely, there’s something that’s still called Ethernet, even though it’s changed so much over the decades that it barely resembles the original idea from the 1970s.
When Amin Vahdat, Google’s chief AI and infrastructure technologist, spoke at the Hot Interconnects conference last summer, he said networks must change to support the hunger for computing power demanded by modern AI applications. According to him, many things have to be relearned, but also forgotten.
Today, Ethernet has to cover an almost incredibly wide range of needs. The same network foundation is used for slow communication over long distances, but also for AI systems that require huge throughput, minimal delay and as little jitter as possible. It is this adaptability that explains why Ethernet did not end up like many previous network standards.
The original philosophy of Ethernet was quite different. When Bob Metcalfe and David Boggs of the Xerox Palo Alto Research Center described Ethernet in a 1976 paper, they were starting from a chaotic environment. Anyone could send data at any time, packet collisions were expected, and the protocol used a random wait before retrying to send.
Ethernet survives because it changes along with the demands of the modern computing age
Such an approach was good enough for early networks, but it did not suit systems that required predictable delays and precise data delivery. That’s why Ethernet started to change very quickly. The coaxial cable was replaced by twisted pairs, passive hubs gave way to switch devices, and packet collisions practically disappeared.
Switch changed the character of Ethernet. Instead of all devices sharing the same chaotic channel, each link could be isolated, and packets temporarily stored and forwarded to their destination. This gave the network greater bandwidth, but also space for more and more computer logic to enter the network devices themselves.
Later, large data centers began to demand even more control. Facebook, Microsoft and other hyperscale operators did not want completely closed switch systems, but software-controlled networks. This is how open approaches such as SONiC and SAI were created, which allow the behavior of switch devices and smart network cards to be adapted to specific needs.
AI has now opened up a new problem. Training large models is not like classic internet traffic. Instead of diverse small streams, AI systems send huge, connected streams of data between thousands of accelerators. When multiple such “elephant flows” end up on the same path, the switch may drop packets that it cannot buffer, and retransmissions waste valuable time.
BenchmarkThat’s why Ultra Ethernet is born, a standard developed specifically for the needs of AI infrastructure. One of the key changes are faster retry signals at the link-layer level, that is, at the level of an individual link segment. The network card can keep a copy of the packet, wait for confirmation, or quickly receive a signal that the packet is lost or damaged. Instead of slow recovery at higher levels, the error is resolved much closer to where it occurred.
Interestingly, such ideas do not have to be confined to AI data centers. Industrial networks also require reliable and fast detection of lost packets, especially in factories, automation and systems where latency can become a serious problem.
At the same time, Ethernet is expanding in other directions. Power over Ethernet allows both data and power to be transmitted over the same cable, which is practical for cameras, sensors and devices in factories, warehouses and shops. With support in SONiC systems, the switch can not only power the device, but also check whether it is really a legitimate device and not something that someone has connected without authorization.
A different idea emerges in AI networks: maybe every lost packet doesn’t always have to be repeated. For machine learning, some loss can be acceptable if the error is distributed so that it does not significantly affect the accuracy of the model. Experiments with the OptiNIC software showed an acceleration of the training of smaller models by more than 30 percent.
Another direction of development is smarter switch devices that not only forward data, but take over part of the computing work. If much of the AI traffic comes down to vector addition and reduction, some of that work can be done within the network infrastructure itself. This reduces the amount of data that has to travel throughout the network.
Ethernet survived because it was never trapped in its first form. From the chaotic office network to switch systems, data centers, PoE devices, cars, industry and AI infrastructure, it has constantly embraced new layers, rules and forms of computing.
That is why today it is not only a question of how Ethernet survived for 50 years, but whether what we call Ethernet today is still the same technology. It may not be in the original sense, but this ability to change without changing its name is the reason why it continues to connect a large part of the digital world, E+T reports.