The past few years have seen fascinating developments in the evolution of the InfiniBand and Ethernet protocols, and in particular, a convergence of the two at the physical link level, if not also arguably at the market level. Intellectual property issues, and most prominently patent rights, are typically intertwined with technological developments. The Ethernet-InfiniBand convergence is no exception, and some of the potential patent implications merit careful consideration by developers, vendors, and users of these interconnect solutions.
InfiniBand was designed to be a universal I/O interconnect, conspicuously not backward compatible with Ethernet and thus free from then-existing Ethernet structures. InfiniBand was meant to be used as a switching fabric to connect both peripherals and networks. For various reasons, however, that purpose was not completely realized, and InfiniBand found its niche in high performance computing because of its low latency. The hallmark of InfiniBand, in many ways, is remote direct memory access, which allows data to be written from user space memory in one server directly into the user space memory of another server, thereby circumventing CPU and network stack overhead that slows down a data transfer.
Ethernet, on the other hand, traditionally has been the cheaper carrier, but to the detriment of relative performance. Ethernet has been more widely adopted and broadly implemented than InfiniBand, offering a greater infrastructure and customer base. Yet in the past decade or so, the gap between Ethernet and InfiniBand has been closing on at least two critical levels, and each implicates important patent issues.
First, the functionalities of InfiniBand and Ethernet have blended together to a significant extent. Some manufacturers have developed and are selling ASICs that support either InfiniBand or Ethernet protocols. Further, RDMA now can also be provided using RDMA over Converged Ethernet and internet Wide Area RDMA Protocol. And even in that Ethernet-supported RDMA sphere, especially in the 40GbE adapter market, many would argue that RoCE has dominated iWARP and has proven to be the solution with staying power. Interestingly, RoCE is supported by the InfiniBand Trade Association and has its own specification (an annex to the IBTA standard), while iWARP is similarly supported by the Internet Engineering Task Force. By borrowing RDMA and reducing overhead, Ethernet has caught up to InfiniBand in performance, resulting in 100GbE and higher products being available.
This development has resulted in a flurry of patent applications to cover invention methods, network devices, and systems in RDMA over Ethernet networks, such as in the acceleration of transmission of packet sequences and in optical transceivers that support both InfiniBand and Ethernet protocols. Many of these applications are ostensibly aimed at reading upon the pertinent standards or where industry participants believe the next iterations of the standards are headed. However, the borrowing of one known functionality to a known protocol may make it more difficult to convincingly articulate the patentability of an invention, including its novelty or non-obviousness to a person of ordinary skill in the art, especially to the extent industry pundits contemplated RDMA over Ethernet was the logical extension of Ethernet. In recent years, both the U.S. Patent and Trademark Office and the federal courts have taken a stricter view of the requirements of patentability, including whether something is in fact patentable subject matter and whether it really would not have been an obvious combination or incremental innovation over existing art. The same concerns are likely to pervade inventors’ efforts to protect under patent law novel gateways or other hardware or software tools to bridge InfiniBand and Ethernet networks, another way the gap between the two protocols is continuing to narrow.
Second, the InfiniBand and Ethernet protocols are converging in their customer and application base. About half of the machines on the Top 500 supercomputer list use an InfiniBand network, and InfiniBand provides more computing power than any other interconnect in the installations on that list. The dominance of InfiniBand in the HPC tranche is clear. The larger customer base outside HPC, such as financial services firms, have not followed suit. Nevertheless, InfiniBand has slowly expanded in recent years to other applications, including clustered databases and cloud storage, because of the latency needs in those applications. Meanwhile, the incorporation of RDMA over Ethernet and the resulting ability of Ethernet to offer 40Gbps performance, although not as fast as 56Gbps or 100Gbps InfiniBand (or the even faster speeds being developed), have proven a satisfactory interconnect solution for a growing customer base.
These market trends invoke questions about the value of patents covering RDMA functionality. Because patents are, at their core, a legal monopoly for a determined time, they of course have some inherent blocking value. But given the different ends of the sandbox Ethernet and InfiniBand still largely play in, proving competition and successfully obtaining an injunction may be difficult in typical situations. Determining the monetary value of patents is equally difficult, but certain observations warrant attention. Patents that are necessary to implement a standard must generally be licensed on the same terms to everybody. Members of the IBTA, for example, must agree to license patents essential to implement either InfiniBand or RoCE specifications on reasonable and nondiscriminatory terms, so the cap on the licensing dollar value of such patents is determined by the number of licensees. In addition, since RDMA is at the heart of InfiniBand and now RoCE and iWARP, the value in owning patents covering inventions necessary to RDMA functionality may also be found in their potential as prior art to new patents or applications for RDMA over Ethernet interconnect solutions.
Despite the convergence between InfiniBand and Ethernet, some predict that InfiniBand supporters will remain loyal, while others predict that a shift to software-defined networking and other factors favor the ultimate dominance of Ethernet, while others still believe that an alternative interconnect scheme to both – such as PCI (News - Alert) (News - Alert) Express switching – will prevail in the market. In this uncertain technological landscape, patent law may be one of the few compasses available, for only those dedicated to protecting their patent rights and respecting those of others will navigate evolving interconnect technology unscathed.
In this uncertain technological landscape, patent law may be one of the few compasses available, for only those dedicated to protecting their patent rights and respecting those of others will navigate evolving interconnect technology unscathed.
Edited by Alicia Young