Intel clearly insists on being the one to specify the next generation of IO that supplants the veritable PCI data bus when it finally sputters to an end. A couple of years ago, most of us probably thought NGIO, which became Infiniband, certainly would replace the PCI bus on servers, and then migrate to PCs as happens always with Intel-Architecture computers. But alas, out of the blue, Intel has decreed yet another next generation IO technology for PCs, obviously because Infiniband is not deemed economical outside the server space.
Yet, 3GIO not only is eerily remindful of the NGIO namesake, but early indications suggest it will share another distinctive characteristic with NGIO and therefore Infiniband; a commitment to serial processing. Indeed, one wonders if Intel intends just to dress NGIO in slightly different clothing (say a stripped down version) and offer it up as (1) adequate for the long-haul and (2) neatly compatible with Infiniband in important ways. If so, Intel must be bothering with this solely to prevent any proposed alternative from gaining critical mass while Infiniband matures, and knowing that over time any 3GIO standards group will restore features stripped from Infiniband.
It seems to me that Infiniband really is the direction toward which Intel wants PC architecture to gravitate. This would make for a unified computing IO architecture that has to be attractive to the architect-in-chief of the PC and now also the server. In the years it will take to get Infiniband ensconced in servers, and then to begin migrating to PCs, surely the cost-prohibitive peripheral intelligence dissipates as an issue on the PC. While today the satisfaction of this requirement requires an expensive IOP, tomorrow the IOP soft core could be grafted cheaply onto SoCs. Alternatively, so-called fixed-function peripheral devices could implement a reduced level of intelligence in hardware (ASIC or a PLD) that would satisfy needs of most PC peripherals. Indeed, even reconfigurable computing might play an interesting and cost-effective role in Infiniband peripherals.
It must be dismaying to Intel, therefore, that many of the important reasons for upgrading servers to Infiniband are no longer pressing like they were a couple of years ago. Recall that Infiniband extends the reach of servers to peripherals and permits up to the equivalent of 65,000 slots to attach devices. However, the inexorable shift toward server blades and network-attached devices obviates both of these needs. The modern server needs only a LAN connection, preferably some kind of a LAN switch or multiple NICs. This suggests the Infiniband ramp will be slower than thought before, and may even struggle to reach critical mass. Coming to the rescue, GbE (with 10GbE looming around the corner) benefits dramatically from intelligent NICs that can offload the TCP/IP stack and improve the necessary data flow to the CPU. Should iNICs become a de facto standard in servers and PCs alike, as it should, then, surprise, it also accommodates Infiniband’s messaging virtually automatically. This is because iLANs have the intelligence capability and the I2O smarts to talk Infiniband messaging.
In the final analysis, the main risk to Intel is not so much what form the next generation IO takes on the PC, but whether there will even be a next generation. The ISA bus was defined at the outset of the PC by IBM and became a de facto standard overnight. The second generation PCI bus grew out of years of bus wars, starting with IBM’s MCA, which was countered by EISA (led by Compaq) and other buses. This war was settled by the late arrival of PCI offered up as an open standard by Intel, by then the architectural leader of the PC paradigm. The so-called third, or next, generation started to bubble years ago, and has continued to bubble and perhaps even boil, but with no resolution in sight. The problem seems to be that to support the PC paradigm for a full generation of ten or so years, the solution has to be remarkably advanced, anticipating future needs and directions of computer architecture as well as easily accommodating a massive amount of existing peripheral devices and technologies. One elegant technology, Infiniband, can’t get legs on the PC apparently because the very elegance probably needed in a few years makes it cost-prohibitive today. Consequently, the shift to a third generation is making the disorderly shift to the second generation look orchestrated in comparison.
If Intel can’t establish a de facto 3GIO, I can’t imagine the PC paradigm would long survive a dismemberment of its fundamental architecture that certainly would obtain. We know the paradigm peaked in 2000, as forecasted for years on this thread. The last thing the paradigm needs is to suffer an architectural splintering along the critical IO fault line.
I have to say, I think Intel made a error in the way they introduced NGIO, one from which they may yet recover, but then again, as discussed above, maybe not. I believe the Intel Server Group, which in 1998 created NGIO and also inherited I2O, decided to downplay I2O for fear of it being a competitive threat to NGIO. In fact, as I argued to the Intel decision makers six months after they made that fateful decision, branding I2O to the point of achieving network effects would be the best way possible to lay the groundwork for NGIO, which at the time of my conversation was over a year away from silicon. Now we see that NGIO (Infiniband) can’t become the 3GIO on the PC because peripherals are too dumb, and Infiniband will continue to struggle on servers because the main problems it solves are evaporating. The irony is that I2O may prove to be the saving grace for Infiniband as well as maintaining the integrity of Intel’s PC paradigm.
All this has and will continue to affect WIND. Had Intel branded I2O, I2O revenues would have ramped many times faster than they have, and WIND would have been prized by the Street. Even so, I2O has been good and highly profitable for WIND, and as always the signs for the future with I2O remain remarkably positive, as illustrated by the opportunity TINA represents and Intel’s public comments about shipping millions of IOPs.
Allen |
