Ethernet in the Backplane
Backplane applications are effectively chip-to-chip transfers, just over longer traces and across connectors (see Figure 1). Currently, the inefficiencies of Ethernet make it suboptimal for these applications, particularly in terms of reliability, effective throughput, responsiveness, and overall latency. Reliability is the Achilles’ heel of Ethernet. In order to reduce system cost, many developers use a single fabric to transport both application data and control plane traffic. Ethernet’s besteffort service, however, is insufficient to transport control plane traffic in a reliable fashion. For this reason, TCP/IP is required to compensate for dropped packets, leading to additional latency and implementation complexity. Serial RapidIO, on the other hand, guarantees packet delivery as part of the base specification and therefore is robust enough as a system fabric to carry control plane traffic reliably. As applications become more complex and process more data more quickly, effective throughput has become a major bottleneck for Ethernet-based systems.
Many applications have used 10/100 Ethernet over the backplane with success, but only when over-provisioning of at least 4X has been implemented to reduce the frequency of dropped packets. As these applications increase in performance, however, Ethernet quickly hits its capacity limits. Serial RapidIO only requires overprovisioning on the order of 20 to 25% in order to cover the bit errors that are an unavoidable part of high-speed serial communications. The difference in actual throughput is staggering. Consider an industrial application transferring data on the order of 10 Gbps. With over-provisioning, Ethernet maxes out at a theoretical best throughput of 2.5 Gbps, while Serial RapidIO exceeds 7.5 Gbps. Further increasing effective throughput is Serial RapidIO’s lower overhead (14-20 bytes compared to Ethernet’s 38 bytes) and higher efficiency at the smaller payload sizes common to backplane applications (see Table). Serial RapidIO also has the great strength of providing link-level flow control. This means that errors are handled in hardware at the MAC/link layer, minimizing latency and improving system responsiveness to errors. Ethernet, on the other hand, handles errors at layers 3 and 4, which means they are taken care of in a software protocol stack. As a result, error recovery introduces high overhead and latency, two factors that developers struggle to minimize over high-speed backplanes.