What Intel's Habana Labs Deal Could Mean For Its Nervana AI Chips

'It's hard to imagine that they would keep both product lines alive,' one analyst says of how Intel's Habana Labs acquisition could impact its Nervana AI chip plans, though not everyone agrees.

A month after Intel made a big splash about its new Nervana neural network processors, the semiconductor giant announced it has acquired Habana Labs, a startup making artificial intelligence chips that could potentially replace Nervana — or at least overshadow Intel's new products.

That's according to some of the analysts who spoke with CRN after the Santa Clara, Calif.-based company announced the blockbuster $2 billion deal on Monday, which will give an important boost to Intel's portfolio of processors and accelerators for AI workloads.

[Related: Intel Unveils 7nm Xe Ponte Vecchio GPU For HPC, AI Workloads]

The potential wrinkle in Intel's acquisition of the Israeli startup is that Habana's products, the Gaudi processor for training and the Goya processor for inference, could replace or overshadow the Nervana neural network processors for training and inference that the company has been developing since it acquired the namesake company, Nervana Systems, for a reported $408 million in 2016, analysts said.

"It's hard to imagine that they would keep both product lines alive," said Karl Freund, who leads consulting for high-performance computing and deep learning at analyst firm Moor Insights & Strategy.

Freund isn't alone in thinking that the Habana deal could signal a change in Intel's plans for Nervana: "It's going to be an either or situation unless they find a reason to keep both roadmaps alive," Kevin Krewell, principal analyst at Tirias Research, said.

But some see room for both product lines. That includes Alan Priestley, a vice president and analyst at Gartner who thinks the Habana acquisition is simply a continuation of Intel's strategy to provide a diverse range of hardware architectures for a variety of workload needs.

"They need to have flexibility, because as this market evolves, there's no one solution fits all," he said.

When asked how the Habana acquisition will affect Intel's forthcoming Nervana processors — which are set to launch next year — an Intel spokesperson said the chipmaker is taking time to evaluate, but there are no immediate changes to the company's product plans.

"We plan to take full advantage of our combined IP and expertise to deliver unmatched computing performance and efficiency for AI workloads in the data center," the Intel spokesperson said in an emailed statement. "We’re not making roadmap decisions on day one. We’ll take time to assess this combination with input from our customers."

Why Intel Is Paying $2B For Habana Labs

What's at stake for Intel is the chance to take away momentum from rival Nvidia and become the dominant provider in the AI silicon market, which Intel said will grow to $25 billion by 2024.

"Intel is very interested in countering Nvidia in the AI space, so this is a great way of getting a product to market faster," Krewell said.

But even then, if Intel is already in the process of launching its Nervana deep learning processors, why did the chipmaker drop $2 billion on a company developing similar capabilities?

Daniel Newman, principal analyst at Futurum Research, said it's about Intel's ability to continue to demonstrate leadership and progress as it deals with a swath of challenges, from its CPU shortage and 10-nanometer product delays to the increased competition it's facing on multiple fronts, whether it's AMD on the x86 CPU side or large customers developing their own silicon, like Amazon.

"It’s the first time in its history that Intel is not necessarily the market-dominant player in an area," he said. "At this point, being among the dominant player for AI chips is of critical importance to Intel."

Newman said while it's possible for Nervana to continue as its own product line, the Habana deal does create some uncertainty for its future. The analyst offered two other possibilities for Nervana: it gets scrapped or Intel uses some of the product's technology and applies it to Habana's work.

What's clear to Newman is that Intel saw something superior in Habana's Gaudi and Goya processors. Because Intel's venture capital arm, Intel Capital, had been an investor in Habana prior to the acquisition, he said, the chipmaker was likely paying close attention to the startup's progress.

"I think Intel immediately saw it as more competitive and capable with inference solutions and more competitive in the training space as well," he said.

In other words, Intel didn't necessarily want to settle with its existing portfolio if there was a better solution in the market that the semiconductor giant already had access to.

"It's still such an early stage [in the AI market] that there's no one perfect answer on how to build good inference and training chips," Krewell, of Tirias Research, said. "This is an opportunity for Intel to hedge their bets."

Two Ways Habana's Chips Stand Out

Freund, of Moor Insights & Strategy, said cloud service providers, in doing early evaluations, likely voiced a preference for Habana's Gaudi and Goya processors over Intel's Nervana NNP-T and NNP-I processors for deep learning training and inference, respectively.

"You can imagine a scene where a company came back to Intel and said, 'Nervana is really nice, but this Habana chip is better,'" he said.

To date, Intel has named Facebook and Chinese tech giant Baidu as early supporters of its Nervana processors. But Facebook has also given lip service to Habana's Gaudi and Goya processors, both of which are being sampled by select customers this year.

"Habana has made some really good progress with hyperscalers," Krewell said. "They were showing up in a number of applications. Facebook was one."

One of the things that differentiates Habana's chips from Intel's Nervana chips, Freund said, is the way data is shared across multiple compute nodes. For Habana's Gaudi chip, this comes in the form of support for Remote Direct Memory Access, or RDMA, a capability currently not seen in other AI chips that allows one computer to access memory from another without taxing the CPU.

"RDMA is extremely efficient for you to share data across a network or share data across a fabric of accelerators," Freund said.

Tied to this capability is Gaudi's use of standard 100GbE Ethernet, which makes Habana's training chip stand out in contrast to Nervana's proprietary interconnect, according to Freund. By integrating RDMA over Ethernet into Habana's chips, he said, the products could offer a more affordable and faster alternative over current options available in the market.

Newman, of Futurum Research, agreed, saying that Intel will be able to scale Habana's chips more efficiently because of their open fabric.

"It's not about the single chip," he said. "It's about the aggregating, of being able to combine the nodes, and this is going to provide a ton of horsepower."

Habana Labs Continues Intel's 'XPU' Strategy

Outside of Intel's new Deep Learning Boost feature for Xeon processors and its upcoming Xe GPU architecture, most of the chipmaker's AI portfolio has come from acquisitions over the last few years, fueling the company's vision of providing a heterogeneous suite of hardware.

"They continue to look to the outside for what succeeds and what doesn’t and they're using their deep wallets to spend when they need to," Krewell, of Tirias Research, said.

This strategy of providing CPUs, FPGAs, GPUs and AI accelerators like Nervana and Movidius is what Intel has begun publicly referring to as its "XPU platform." The XPU is not a product or architecture itself, but the idea that the evolving nature of computing requires a more diverse set of processors.

Priestley, the analyst at Gartner, said in the AI market, the architecture that's best optimized for computer vision may not be best suited for other types of AI applications, like conversational AI, which is why it's important for Intel to have a variety of options.

"Having multiple options there and the ability to run development in parallel can give you flexibility going forward," he said.

By adding Habana's chips to Intel's portfolio, the chipmaker will have even greater flexibility in giving customers what they need for hardware, which is why Priestley thinks Intel's Nervana chips will stick around.

"The key is going to have the right solutions available for the type of solutions and algorithms being deployed," he said.

But for now, Nervana's future is up in the air as Intel evaluates its product roadmap with customers — who will ultimately decide the product line's fate, according to Krewell.

"It's very possible that the Nervana parts get ditched. Intel is not in the business of keeping it for around just for the sake of 'hey, we spent this much.' There needs to be a market," he said. "If that market is serviced by the Habana chip and the Habana chip is performing better, I certainly think Intel will ditch their own internal part."