What effect will commoditization have on high-performance computing?

Print Friendly, PDF & Email

Technology acquisition decisions in any well-run company are based on the return-on-investment that a purchase will produce. This is what differentiates business customers from consumers, who buy to improve the quality of life. Thus, consumers will gladly pay for luxury items; businesses will not. For this reason, many corporations will take “adequate” over “superior.”

One fundamental premise every technology company must hold when selling to the business market is that cost is king. Therefore, all technology tends to go commodity, or simply disappears.

Consider the x86 processor. While the Cell and the UltraSPARC T1 will have their customers, AMD and Intel (not Itanium, though) can truly claim a massive market presence simply by having a lower total cost of ownership. A customer who needs specialized processing capabilities is usually best served by having x86 CPUs supplemented with a co-processor, whether a broadly available chip or case-specific FPGA.

Another example is Ethernet. Despite all of the innovation in networks previously detailing here, Ethernet still reins supreme. Indeed, Myricom’s newest product line builds on the company’s previous successes to enter the Ethernet market with high bandwidth, low latency, and low utilization. (This move garnered some praise previously on this blog.) Furthermore, the InfiniBand crowd is moving towards iWARP, going so far as the rename the OpenIB Alliance to the OpenFabrics Alliance. Also, many storage customers are trying to move away from Fibre Channel onto general-purpose networks with iSER.

Commoditization does not simply exist in hardware. Microsoft understands this and has made the PC a commodity device simply by licensing its OS to others for distribution. Again though, the total cost of ownership comes into play. Microsoft’s real advantage is in providing excellent programming tools. Likewise, Linux would never have unseated Unix had the former not copied many of the programming aspects from the latter, including POSIX compliance.

Regarding commodity programming tools for networks, Sockets has been a champion regardless of the numerous complaints (co-processors really helped here). As previously detailed, the newer DAPL interface stands to gain wide-spread usage simply by being available for most networks. This requirement is primarily why MPI has achieved so much success, again despite the complaints. It therefore appears than a network vendor would do well to supply DAPL and Sockets, whereas a vendor of parallel technical codes should write for MPI.

In sum, commoditization of both hardware and software is an inevitability. To a degrees this is a shame, as commoditization is just as destructive to innovation as monopolization. Most innovations in networks, processors, operating systems, and programming libraries, among others, have simply not caught on. Reasons range from initial price to invalidation of legacy systems, but almost all are simply a matter of barrier to adoption; it is too difficult, cost-wise, for most customers to accept new technology.

Given this, a technology entrepreneur had better improve on what already exists. (See the discussion of supercomputer vendors for an example.) To borrow the old expression from scientists, one must stand on the shoulders of a giant. Only then can innovation take root in the market.