The chip market has been stable for decades. Mega-producers such as Intel and AMD have dominated. Niche and custom markets were served by a host of other providers. And a few select companies like IBM have continued to develop their own chips.
IBM’s deep pockets mean it doesn’t just produce microprocessors – it is a major innovator. It just announced a 2-nanometer (nm) chip, compared to the usual range of 7-10 nanometers and in some cases, 5 nanometers.
But IBM is being joined by an elite group of cloud providers who are building custom chips designed for their own massive data centers. Google, Facebook, Amazon and Microsoft are all getting in on the act.
Also Read: IBM’s 2nm Breakthrough: Implications for Chip and Server Makers
Custom servers came first
In many ways, what is happening with microprocessors is the extension of a trend that has been going on for a decade or more.
“The hyperscalers like Amazon, Facebook, Google and Microsoft Azure have shown a trend over the past decade of developing their own hardware directly, or via acquisitions, to support general server offloads as well as optimize network, security, encryption, storage I/O, graphics, and other functions,” said Greg Schulz, an analyst with Server and StorageIO Group.
First, these companies designed their own massive data centers and included customized cooling arrangements and introduced solar panels and other renewable resources for power.
Next, they began tailoring servers to their own specifications. They didn’t want the plain vanilla servers provided for the general server market. They wanted servers which were all compute and little or no storage. This enabled them to erect row after row of highly dense compute resources to power their data centers. They also arranged them so they could slide out in seconds when they failed, to be replaced with another blade.
“It’s all about economies of scale, reducing cost, physical and power footprint, component count, and getting more productive work done in a cubic meter or centimeter,” said Schulz. “Not everything is the same, so if you don’t need a lot of memory, or I/O, or compute, you can set up different modules and platforms optimized for a specific workload without carrying the extra physical hardware packaging to support flexibility.”
Let’s make a deal
That trend has continued to this day, and now chip customization seems a logical extension. As in their previous forays into data center and computer design, it is all about optimization of workloads and overall performance, power and cooling efficiency, latency, density, and of course, cost. Don’t think for a minute that Amazon is planning to pay top dollar for customized chips like most people do. Similar to how the online retailer can produce a branded version to undercut a best seller, Amazon and its high-tech pals are likely to cut deals with Asian-based chip producers at heavy discounts.
“Some of the work being done is in collaboration with the Open Compute Project, and the different players are sharing varying amounts of detail on what they are doing,” said Schulz. “We have seen the hyperscalers evolve from using their custom-designed application specific integrated circuits (ASICs) and field programmable gate arrays (FPGAs) as well as hypervisors to more full-fledged built-to-order chips with custom components for servers, adapters and other devices.”
Google, too, has been especially active in tailoring compute resources to its needs, all the way down to the component level. It has been working closely with partners to produce its own solid-state drives (SSDs), hard drives, networking switches and network interface cards (NICs).
In many ways, these companies drive the market to go in directions that suit their needs. Google already has its own customized chips such as the Tensor Processing Unit (TPU) for machine learning workloads, Video Processing Units (VPUs) to improve video distribution, and even open-source chips aimed at enhanced security.
“Custom chips are one way to boost performance and efficiency now that Moore’s Law no longer provides rapid improvements for everyone,” said Amin Vahdat, Google Fellow and vice president of systems infrastructure.
Now comes system-on-a-chip (SoC) designs. Instead of lots of different components on a motherboard, the idea is to integrate them all within a microchip package. The goal, as always, is to lower latency, power consumption and cost while obtaining higher performance. This pits Google against AMD, ARM, and others who already produce SoCs. Other hyperscalers like Facebook and Microsoft are getting in on the SoC act, too.
“Instead of integrating components on a motherboard where they are separated by inches of wires, we are turning to SoC designs where multiple functions sit on the same chip, or on multiple chips inside one package,” said Vahdat. “In other words, the SoC is the new motherboard.”
Schulz believes the hyperscalers will succeed, as they have done in similar forays in the past.
“Given their size, ecosystem, scale of deployment, they can succeed as we have already seen with AWS and its Nitro project (hardware and software), along with what Microsoft, Google and others are doing,” said Schulz. “At this point, they may not go as far as designing CPU instruction chipset-level products and directly competing with Intel, AMD, and ARM.”
Intel and AMD response
In light of this new development, how will Intel and AMD respond? ServerWatch reached out to Intel and AMD, but neither commented.
Schulz said Intel not only needs to deal with the actions of the hyperscalers, it also needs to respond to recent innovations from AMD, ARM and Nvidia — the last two of which are merging in a blockbuster $40 billion deal.
“Intel could eventually position itself to be the fab of choice for the hyperscalers while also supplying CPU core and other technology to be used under license to these tech giants,” said Schulz. “Whatever happens, Intel and AMD will need to add more functionality, become more flexible, and remove costs and complexity to be competitive.”
However, he doesn’t think the x86 chip market is going away anytime soon. Over time, a greater percentage of GPU, FPGA, ASIC, and other specialized microchips will be added to custom motherboards, adapters, and mezzanine cards. Likewise, Schulz predicts that we will see more system integration on motherboards for the Internet of Things (IoT) and other small platforms, as well as mobile devices.
Given Amazon’s retail reach, are we likely to see Amazon branded microchips? Schulz doesn’t think so, at least in the near term.
“Initially, any consumption of custom chips will be purely internal,” he predicted. “However, we will see custom AWS technology deployed into data centers.”
Given Amazon’s seeming ability to venture into any market, that one in particular bears watching.