Facebook’s Plan To Design Its Own AI Chips Has Set Alarm Bells Ringing For Qualcomm, Nvidia And Intel

...

Posted by Matt Walker  on May 30, 2018

[Ed. note: see Sept. 2018 publication, Webscale Network Operators: 3Q18 Market Landscape]

Rumors about Facebook’s likely entry in the hardware segment were put to rest after the company posted job openings for chip designers. Facebook was looking for candidates specialized in architecting and designing ASIC and FPGA chips, to help build “custom solutions targeted at multiple verticals including AI/ML, compression, and video encoding.” Facebook has data center applications on the mind, such as live video content filtering, but the company may also be building chips to support its Oculus virtual reality headset and long-planned smart speaker.

Facebook’s chip plans are risky but may bring increased control over supply chain

Facebook’s recent move to make its own artificial intelligence (AI) chips and get a foothold in the hardware segment is a step in the right direction, as it looks to reduce dependence on chip manufacturers (Intel, Nvidia and Qualcomm) while putting a lid on its costs. However, designing a chip is by no means a simple task and not a core competency at Facebook. Matching the performance and efficiency of Intel, Nvidia and others will be a challenge.

Why would Facebook take the risk? It’s loaded with cash ($41.7B in Dec 2017) and used to making high-stakes tech investments. But the chip market is competitive, and Facebook has substantial buying power – it could certainly rely on the open market. Time to market may improve with self-design, for sure. However, another benefit may be more persuasive: Facebook gets greater control over intellectual property rights and information flow. It likely has a few surprises in store.

Big technology investments needed to support social networking

Facebook’s growth has been driven by acquiring and strengthening complementary services to its social networking business, such as WhatsApp and Instagram. Supporting this growth enticed Facebook to build a huge core network.

Like other webscale providers, Facebook works with contract manufacturers to build custom servers and other gear for their massive data centers. Facebook has played an important industry role in this regard, serving as an early sponsor for both the Open Compute Project (OCP) and Telecom Infrastructure Projects (TCP). Now Facebook is testing the waters in consumer electronics markets, initially with its Oculus virtual reality headsets, a smart speaker to be launched in 2018, and complementary AI software. These efforts have contributed to both high capital spending and R&D expenses at Facebook (Figure 1).

Facebook believes that creating its own custom designed chips will result in better integration of hardware and software, and give it tighter control over the development of the product. One factor behind Facebook’s chip push is an interest in running AI algorithms in-house, to avoid sharing with third-party vendors like Intel or Qualcomm. Not every webscale company can do this, but Facebook is positioned better than most due to its deep capex budget & its pioneering work at the OCP, TIP, and other groups.

Working in Facebook’s favor is its recent partnership with Intel to manufacture its own AI processor last year.

Impact of Facebook’s entry into the semiconductor space on the big chipmakers

In the past two years, there has been growing tension between the tech players building cloud networks and the vendors they rely on, mainly Intel, Qualcomm, and Nvidia. Historically the largest cloud builders, “webscale network operators” (WNOs) in our terminology, have heavily relied on Intel’s microprocessors and Nvidia’s GPUs to power their data centers (Figure 2).

However, Facebook is not the only WNO to look at building its own chip. Many webscale providers are starting to look in-house to build custom AI chips to reduce costs and improve on efficiency. For instance, Google developed an AI chip, Cloud Tensor Processing Unit (TPU), two years back, to boost its AI workloads. Google released a latest version this month, indicating that Nvidia’s dominance as a supplier of AI chips could soon be in jeopardy. Similarly, Apple plans to build its own chips for its Mac desktops by 2020, thus reducing its dependence on Intel. Amazon is building its own custom hardware to improve its Alexa enabled devices. Microsoft has launched a new cloud service for image-recognition projects powered by its FPGA technology, codenamed “Project Brainwave”.  This will rely on Intel Stratix 10 chips and support a neural network based on the ResNet-50 architecture. Microsoft claims that this new technology will be capable of handling AI tasks rapidly enough to be used for real-time jobs and at a reduced cost in comparison to the graphics chips (e.g. NVIDIA) used in machine learning tasks.

As webscale tech players build more of their own chips, traditional chip developers are getting nervous.

One company affected in a big way by Facebook’s move is Qualcomm. Facebook’s new chips may be used to power its VR headset, Oculus Go, which currently runs on a Qualcomm Snapdragon 821 chip. This could be a huge blow for Qualcomm. And it comes at a time when Qualcomm is already struggling after its legal battle with Apple, an attempted acquisition from Broadcom, and a still-pending merger with NXP.

The chip vendor’s fears are not just theoretical. Webscale players have already had an impact on the supply chain, hurting server vendors like IBM and HPE in past years by going to white box/contract manufacturing. Now they’re big enough to design their own chips. That cuts out the middleman, avoids having to share secret IP, maybe speeds time to market, and may result in some proprietary advances.

Facebook will have to win back faith amidst data privacy scandal

 While Facebook engineers will continue to find ways to make their network cheaper and smarter, the company faces more complex challenges in the area of data privacy & public perceptions.

Facebook’s privacy practices have come under global scrutiny in recent months, due to the recent Cambridge Analytica and Android call data scandals, not helped by a photo tagging-related lawsuit. As a consumer-facing brand with plans to build its own IoT hardware, a lot is at stake. The company needs to build trust and improve transparency. It cannot do this while also maximizing ad revenue growth. Not all Facebook executives seem willing to accept this.

Amid all the public outcry, the launch of Facebook’s smart speaker has unsurprisingly been delayed. With recent news that Amazon’s Alexa has some interesting privacy-related glitches, Facebook’s decision is probably best. Now seems like a good time to focus on the basics.

TAGS: Capex, Data centers, Webscale/Hyperscale, Artificial intelligence (AI), R&D, Chips

Subscribe & stay updated

Sign up for MTN Consulting's Blog and receive our latest updates straight to your inbox