There’s a transformation in CPU design taking place just below the surface of the market. In the long run, it could prove every bit as important as the shift from single-core to multi-core processors, or the transformation of the x86 architecture from a me-too product for personal PCs into a market-spanning titan. Intel’s CPUs aren’t just being designed by Intel anymore. Companies like Google, Facebook, and Amazon are now playing a direct role in hardware design.
To understand the significance of this change, we need to revisit how CPU design has evolved over the past few decades.
In the mid-1980s, Intel rapidly devoured the desktop market. In the mid-1990s, it ate the datacentre. Obviously these market gains were driven by close customer partnerships, but the late 1990s saw the debut of RDRAM, Itanium, and the P4 (Netburst) architecture – technologies arguably designed to benefit Intel’s bottom line far more than the customers that purchased them.
Competition from AMD, concerted pushback from the DRAM industry, and the failure of Dennard scaling sank these initiatives and forced Intel to consider new alternatives in the form of multi-core products and what we’ve previously called “More than Moore” scaling. More-than-Moore (MtM) is defined by the ITRS (International Technology Roadmap for Semiconductors) as addressing “an emerging category of devices that incorporate functionalities that do not necessarily scale according to Moore’s Law, but provide additional value to the end customer in different ways.”
Working with major datacentre partners is becoming a critical way to find and identify new opportunities that go beyond simply scaling parts to lower power consumption and hitting higher performance targets. It’s not that HP, Dell, and IBM are no longer important to the server business, but rather that Facebook, Amazon, and Google now do huge business volumes as direct customers with custom projects and specific needs.
The exact degree of customisation that Intel implements for these companies is still unclear. It’s easy to imagine Intel offering high-clock SKUs with correspondingly higher TDPs, or working to develop low power products, but a recent report indicates that the collaboration goes much deeper. According to Frank Frankovsky, head designer of hardware for Facebook, the company works with Intel quite closely.
“The more insight we can give our technology suppliers about what makes our software work the best, the more we’re able to influence their design roadmaps,” Frankovsky says. “We do influence their roadmaps – way upstream – but then they are able to bring these changes out to all of their customers.”
This could be seen as an outgrowth of the “dark silicon” phenomenon. One of the problems with the phrase “Moore’s law” is that it’s historically been used in two very different ways. The original version of Moore’s law states that transistor density doubles every 18-24 months. Later, this was expanded to include the phenomenon known as Dennard scaling. Dennard scaling was the observation that as transistors got smaller, the power used by each transistor shrank.
Moore’s law – the increasing density of transistors over time – still holds true. Dennard scaling doesn’t. As a result, it’s increasingly difficult to turn all the transistors in a CPU on at the same time while remaining within a given thermal envelope. This is a problem that’s only getting worse over time.
There’s no single solution at this point, but low density specialised silicon for implementing certain capabilities in hardware is one method of addressing the problem. Intel’s NTV (Near Threshold Voltage) research and the shift towards heterogeneous computing are both approaches that focus on utilising dark or dim silicon to improve performance without sabotaging power consumption.
AMD has said it also handles some semi-custom work, but I’m betting the company’s hands are generally full of PS4, Xbox, and Wii U-related work at this juncture. TSMC and GlobalFoundries will happily build Facebook a custom core, but don’t maintain full design houses to create such a product from scratch. For now, this is a fairly potent Intel advantage. Looking at the list of new features in Intel’s 22nm Atom, it’s impossible to tell which ones are the result of close collaboration with Intel’s new partners, but we’d bet at least a few of them are.
Facebook’s comments imply that while it’s working with Intel to implement new capabilities, not all the features are explicitly exclusive. Intel, in other words, might have added AES-NI instructions to Atom because certain customers wanted those capabilities, but it’s scarcely hiding the fact. For now, none of the companies involved want to give specifics. Once you understand the problem, the idea of Intel doing custom work for a new type of customer makes far more sense.
The standard server OEMs still have an important role to play in distributing general computing hardware, but for Intel, working with Facebook, Google, or Amazon gives the company long-term visibility into its customers’ needs. It also helps reduce the chance of a sudden ARM coup. While we haven’t seen the text of any agreement between Intel and these new customers, it’s safe to assume that the major companies are contributing to the customisation work and have committed to using the final product.