This article comes from optical communication depth: Why, How, and What of Ethernet in AI computing power for more information, see RoCE replaces IB: why is it IB before and RoCE now?".
Extending on traditional Ethernet, the super Ethernet Alliance gathers head players. Super Ethernet Alliance ( UEC) by Linux sponsored by the Foundation and its Joint Development Foundation, the goal is to surpass the existing Ethernet functions RDMA and RoCE and so on HPC and AI the high-performance, distributed, and lossless transport layer of computing. Its initial members include AMD, Arista, Botong, Cisco, Eviden, HPE, Intel, Meta and Microsoft.
As2024 year3 month19 day, UEC added45 new member and published UEC specification1.0 overview White Paper, briefly UEC the specification can realize eight functions and ultra-Ethernet transmission ( UET).
Botong is the world's leading wired and wireless communication half conductor company, has been deeply engaged in the industry at present60 over the years, it has profound technical accumulation and rich product portfolio. In RoCE field, the company from controller, adapter, NIC, switch from four aspects, currently there are super30 A variety of related products, recently Botong is based on the fourth generation RoCE launch single port400GbE ethernet adapter N1400GD and single port400GPCIe ethernet NIC P1400GD, mainly applied AI, cloud computing, high-performance computing, and storage network construction.
Avida in NIC and switch direction bureau, although Avida was InfiniBand the main promoters and suppliers, but also continue in RoCE direction layout, launched one after another Spectrum SN4000 and Spectrum SN5000 vswitches, and launched this year IB new products of the same specification Spectrum X800 switch2025 launched in512 port Spectrum UltraX800 VSwitch2026 compared with the annual bandwidth X800 doubled X1600.
2020 since, Meta always committed operations based on RoCE however, it faces consistency challenges in the early stage. To implement RoCE of AI computing applications are implemented, Meta as a founding member, we established the super Ethernet Alliance and actively promoted it. RoCE the deployment. Company usage Arista 7800 and Wedge 400 equal RoCE network Implementation400g interconnection has been successfully applied Llama3 cluster.
RDMA compared with traditional TCP/IP the technology is more in line AI high Concurrency and low latency are preferred. And the previous TCP/IP compared with the hardware and software architecture, RDMA enables the communication system to access directly through the NIC GPU video storage data, the process does not need to go through the operating system or CPU, this high throughput, low latency network communication is very suitable for large-scale parallel AI used in computing clusters.
Currently supported RDMA the network Infiniband, RoCE(RDMA over Converged Ethernet), iWARP, different network features:
Infiniband: specially designed RDMA designed to ensure reliable transmission from the hardware level, the application effect is good, no need do Targeted Design and development, but need IB nic and switch support, high cost
RoCE: based on Ethernet and transport layer UDP protocol design consumes less resources and can use common the Ethernet switch, but need special support RoCE the NIC.
iWARP: based on Ethernet transport layer TCP protocol TCP reliable transmission is achieved. Compared RoCE, in the case of large-scale networking, iWARP A large number TCP connections consume a large amount of memory resources ( RoCE of UDP connection is not required), which requires higher system specifications. Common Ethernet switches can be used, but special support is required. iWARP the NIC.
In AI in the wave of computing power construction, IB is the early local optimal solution, RoCE is a broader optimal solution. In AI calculate at the beginning of the acceleration of construction, high throughput and low latency network requirements need to be supported. RDMA network communication from Avida H series GPU it can also be seen from the continuous shortage of demand that the implementation of computing power quickly, with good quality and quantity in a short period of time is the core demand of all computing power investors. Therefore, Avida's GPU add natural adaptation RDMA of IB network architecture is the optimal solution at that time.
In the long run, Ethernet/RoCE compared IB it has a deeper industrial application foundation in the field of cloud computing and achieves costs. It is also lower. As the technology becomes more and more mature and the inference demand gradually rises, Ethernet will gradually come AI power Dance center.
Source: technical architects Alliance
