Blog Post

Amazon joins other web giants trying to design its own chips

Stay on Top of Enterprise Technology Trends

Get updates impacting your industry from our GigaOm Research Community
Join the Community!

As demand for its cloud computing services continues to grow, sources say Amazon is trying to design its own server chips. Based on a job listing and a series of LinkedIn updates, it looks like it could be eyeing the ARM architecture for those chips.

The online retailer and cloud giant has hired several chip engineers who used to work at Calxeda, the former ARM-based server startup out of Austin, Texas that shut down last year, including the former Calxeda CTO. It also has a few job listings for its Silicon Optimization team based in Austin, Texas that call for microprocessor design expertise, including one for a “CPU Architect / Micro-Architect.”

Amazon declined to comment on its plans; actually, a spokeswoman said, “We don’t comment on rumors or speculation.” Don’t worry, we’ll make sure to ask Amazon’s CTO Werner Vogels (pictured) about these plans at our Structure event in June. However, several sources in the Austin chip community have been discussing the impact of Amazon’s decision to swoop in and recruit engineers after Calxeda shut down in December, curious about whether or not Amazon planned to follow other webscale businesses Google and Facebook in trying to build their own silicon.

Frank Frankovsky of Facebook holding an AppliedMicro ARM board.
Frank Frankovsky of Facebook holding an AppliedMicro ARM board.

A risky bet and unorthodox thinking

Now, a few sources who aren’t authorized to speak publicly about Amazon’s plans say Amazon is indeed going that route. Judging by a search of LinkedIn profiles, Amazon has hired at least six former Calxeda employees with both silicon design, fabric and board design experience. The new Amazon hires came in on in March with job titles that include, Principal Engineer, Silicon Optimization at Amazon Web Services; Hardware Development Engineer; Hardware Design Engineer; Director of Silicon Optimizations at Amazon Web Services; and Manager of Hardware Engineering, Silicon Optimizations at Amazon Web Services.

When pointed to the job listing and Calxeda hires, Patrick Moorhead, president and principal analyst at Moor Insights and Strategy, said that it looks like Amazon is investigating custom chips as opposed to just working with existing chip suppliers to get more highly optimized silicon and boards.

“That listing looks like they are taking an ARM architecture license to roll their own core,” he said. “Why else would you need an architect?”

But taking on silicon design is a big bet, even though it has become more feasible as operating systems and supporting software built around alternative silicon architectures become more visible.

“It is a major investment in the hundreds of millions of dollars to build your own custom core but if you buy hardware like Amazon does, and if you are truly successful at matching your workloads to your processors, then it could pay off,” Moorhead said. “But it’s very risky.”

He added that he doesn’t know if any of those trying to design their own cores are going to be able to compete with Intel on pure performance, but noted that if Amazon were looking at using lower-performance cores in a massively distributed fashion (which would also be a reason to hire the Calxeda engineers who are expert at building out distributed compute fabrics) that could create cost optimizations that could justify the expense of “rolling its own core.”

In the commodity cloud even chips can be squeezed

Speakers: Andrew Feldman - GM and Corporate VP, AMD
Andrew Feldman – GM and Corporate VP, AMD

Despite the challenges, Amazon’s foray into silicon makes sense. As we’ve written before, the costs of designing your own silicon have dropped while the penalties for not cutting as much cost as possible from your hardware operations have risen.

Last November Amazon’s James Hamilton explained the company’s thinking about building its own servers and infrastructure, pointing out that it saved 30 percent on costs.

Like Google and Facebook, Amazon is designing its own servers, and they’re all specialized for the particular service they’re running. Back in the day, Hamilton used to lobby for just having one or two SKUs from a server vendor in order to minimize complexity, but times have changed. Once you master the process, going straight to server manufacturers with custom designs can lop 30 percent off the price right away, not to mention the improved performance and faster turnaround time.

Today, “You’d be stealing from your customers not to optimize your hardware,” he said.

And now with Google and Facebook investigating chip designs optimized for their workloads, Amazon can only do what it can to keep pace with its rivals; especially Google, who will likely be Amazon’s biggest competitor when it comes to offering a commodity cloud platform. If you take IBM’s decision to use its open Power8 architecture in its SoftLayer cloud (Google is also evaluating that architecture), it too, is using custom silicon for its cloud offering. That of course, leaves Microsoft’s plans an open question.

The looming introduction of a 64-bit ARM-based server core (production 64-bit ARM server chips are expected from a variety of vendors later this year) also changes the economics of developing a server chip. While Moorhead believes building your own core is a multihundred million dollar process, Andrew Feldman, the corporate vice president and general manager of Advanced Micro Devices’ server chip business, told me last December that it could be in the tens of millions.

Feldman also said, that for the webscale players, it’s not a question of if they will develop their own chips, but of when. At the time he said he suspects we’ll see custom chips coming out in the next three to five years. However, to make that possibility a firm would have to start designing today.

That’s appears to be exactly what Amazon is doing.

10 Responses to “Amazon joins other web giants trying to design its own chips”

  1. The problem is that custom ARM chips for servers do not make economic sense for anybody. Designing, validating and taking a server chip to production is insanely expensive and you need to do it on a treadmill of at least every 2 years in order to stay remotely competitive. If a custom server chip takes $100M R&D to make (how much money did Calxeda burn through?) and if you are wildly successful and sell 1M chips (you won’t sell that many) then your R&D cost per chip would be $100 before you start paying TSMC the production costs.

    Your competition (Intel) is able to use cores like Haswell and beyond that offer several times the performance of an ARM core, manufacturing costs lower than TSMC, silicon process 2+ years ahead of the competition, world class SW & HW ecosystem enabling, and a 1 year Tick-Tock innovation treadmill.

    AMD could not make money being one of the only two x86 server CPU companies. How could anyone else really expect to make money being 1 or 10 or 20 ARM server CPU companies?

  2. Remember when people thought “Intel is safe in the server market, anyway” when ARM was starting to push in the consumer market?

    Remember Innovator’s Dilemma, folks. Intel can’t hide up-market for long. It’s just a matter of time before ARM chips eat their lunch there, too.

  3. Certainly makes sense for large scale infrastructure providers to build their own custom components. At the scale Amazon, Google and Microsoft are at, even small increases in efficiency will have a big impact on performance and cost, in particular energy usage. We’ve already heard how Amazon has been designing it’s own power infrastructure all the way out to transformers.

  4. Apple’s success here has made everyone else hungry for similar benefits. There should be one warning about this. It is all predicated on Moores Law continuing to fail for the forseeable future. The reason 30% optimization is so appealing is the usual improvement per process node right now has been on the order of 10% less power and 10% more performance (and the nodes are getting farther apart too). If for some reason there was a break through on a new process to make computer chips scale again, then we would see far smaller returns for your money. The optimization is based upon the whole platform changing slowly enough that you can optimize before a new chip type is necessary. With Moores Law and especially Denson scaling wheezing along at such an anemic pace, it is much easier to design chips in 3 years and use them for 3 years. Imagine if the whole platform had to change after 4.5 years. Back in the 90’s we saw 8 fold improvement over a span of 4.5 years. This required a whole new type of data transfer and management to be considered.

  5. realjjj

    “Andrew Feldman, the corporate vice president and general manager of Advanced Micro Devices’ server chip business, told me last December that it could be in the tens of million”

    Don’t think he said that , it seems he was talking about custom chips not custom cores. After all Moorhead is a former AMD VP and he knows the costs too, they were just talking about different things.