The chip market is a house divided. On the smartphone and tablet front, virtually every device you use, manage, and love is powered by tiny system-on-a-chip (SoCs) designs, processors which present a fundamental leap in the old design paradigms of decades past. Meanwhile, servers and PCs, while offering terrific performance, are still holding on to dated design models.
Evolution in the mobile market has been defined by the rise of the IP core, licensed circuit designs. This flexible approach allows companies like Qualcomm to compete -- and win -- in the mobile space against veterans like Intel, which have vastly bigger research and development budgets. The key is teamwork and agility. With the IP core paradigm, licensers like MIPS Tech. (now owned by Imagination Technologies) or ARM Holdings release basic circuit designs, with all the fundamental logic blocks. This takes the basic design load off the implementer, who merely has to customize the chip with extras like I/O (think LTE modems, image processing, and so on).
By comparison, the server and workstation markets still follow a dinosaur of a chip design paradigm. Each OEM goes out and does the entire design process itself, producing monolithic chips. The approach is sort of the digital equivalent of Ayn Rand's objectivism -- in the sense that it is radically individualist.
But this age of radical individualism in the PC market is coming to a close.
IBM has been working closely with ARM since 2011 to develop a co-designed mobile SoC dubbed Cortex-M0, which will put nearly all the components of the smartphone -- RAM, wireless modems, CPU, GPU, power management circuitry, and possibly even the camera sensor, onto a single tightly wired CPU. While no official announcements have been made, it is a pretty safe bet that IBM will look to apply the lessons learned to its server chips.
AMD is following a single route. It's promoting a concept called "heterogeneous computing" and has also announced it will be offering ARM Opterons (64-bit) to the server market in 2014. It took me a while to wrap my brain around what exactly AMD means by
"heterogeneous" because the company is purposefully vague about the term.
But I believe it's not just buzz. Heterogeneous computing could arguably be defined as the practice of mixing co-designed GPU cores, larger CPU cores (think Power6, Ivy Bridge, or PileDriver), and smaller CPU cores (think smartphone ARM Cortex cores). Math-heavy loads will be shuffled to the GPU cores; lightweight virtualized loads can be shuffled to the small CPU cores; branched processing-intensive will be loaded onto the larger CPU cores.
To date, no one player has achieve this triple architecture. But it is coming, as each of these fundamental building blocks -- the graphics core (essentially, a parallel math coprocessor); the lightweight core; and the monolithic core -- each have a valuable and unique role.
Co-designed, customizable, mixed, triple-component core (CD-MTCC) chips will be perfectly suited for power savings and a highly virtualized datacenter. In the datacenter of the future, large hosting firms will be able to work with chipmakers to produce custom-fit cores. For example, there will likely be a subset of servers with more graphics core-heavy chips, capable of serving online gaming clients needs. There will be a subset of servers with lighter CPU core-heavy chips, which will support clients whose service is driven on lightweight transactions (e-commerce, for example). Finally, for web application clients, there will be larger CPU core-heavy chips. Each type of chip will have some of the other core kinds, to support the odd cases where more of the less typical kind of processing is needed.
Ultimately, this approach will kill the traditional monolithic core and conquer the server market. It will also likely necessitate a wholesale transition to RISC instruction sets, to support a modicum of code homogeneity between the three kinds of cores. I expect even Intel, which is already working to merge GPU IP cores with some of its chips, to adopt a similar MTCC SoC approach, even if it alone has the resources to buck the greater overarching co-design push.
RISC server SoCs will pop up in 2014. Expect the next arrival to be mature designs combining heterogeneous RISC CPU core sizes in the 2016 to 2020 timeframe. Users can simply sit back and enjoy the coming improvements, as this new hardware model should decrease costs and improve service. For those in the serving business and hardware front, however, a tough decade is coming up as the industry prepares to make this radical shift.
But the benefits will be great when the work is done.
— Jason Mick is senior news editor at the independent tech news site DailyTech.