What Does It Mean For AI (Artificial Intelligence)?
This week Intel agreed to pay roughly $2 billion for Habana Labs. Based in Israel and founded in 2015, the company is a startup focused on AI (Artificial Intelligence) chips. Keep in mind that Habana has raised a total of $75 million, which is a fairly modest amount for a hardware company (Intel Capital was one of the investors).
According Intel’s executive vice president and general manager of the Data Platforms Group, Navin Shenoy: “This acquisition advances our AI strategy, which is to provide customers with solutions to fit every performance need–from the intelligent edge to the data center. More specifically, Habana turbo-charges our AI offerings for the data center with a high-performance training processor family and a standards-based programming environment to address evolving AI workloads.”
Already, Intel has seen traction with AI chips, bolstered with other acquisitions for companies like Nervana Systems and Movidius. For 2019, the expectation is that revenues will hit abut $3.5 billion for this segment, up 20% on a year-over-year basis. Intel also forecasts that the total addressable market will be greater than $25 billion by 2024 and half of this will be for the datacenter.
Now when it comes to AI, the software side of the industry gets much of the attention. But to allow for the true power of this technology, there needs to be strong innovation with chips as well.
“Satisfying AI workload requirements is a growing challenge for many organizations,” said Mike Leone, who is a senior analyst at ESG. “Traditional compute is simply unable to keep up with the orders of magnitude improvements organizations are looking for in their respective compute infrastructure. And it’s a losing proposition to just keep throwing more and more processing power at the problem. It’s too expensive. It’s too big of a footprint. And it’s too power hungry. We’re seeing an increase in the need for specialized compute to address the different workloads in the AI space, mainly training and inference. Training addresses the algorithm creation process, by feeding a model data so it can learn. Inference refers to the stage where the trained model gets leveraged to make predictions based on new incoming data. Of the two, training is far more resource intensive. And while GPUs, for example, can address both types of workloads, the emergence of specialized compute based on the AI workload—that is, training vs. inference—has emerged and amassed a surprising number of startups looking to add their IP and approach into the mix.”
Mukesh Khare, who is the vice president of IBM’s AI Hardware Research Center, agrees that AI chips are becoming more critical for AI: “Today, AI applications are being executed on systems designed for other, non-AI purposes. The rapid escalation in AI deployments is straining the capabilities of these systems, and expected overall improvements in general-purpose computing systems cannot keep up with this escalation in demand. For example, the compute needed for AI training is doubling every 3.5 months. To address this AI compute demand growth and opportunity, heterogeneous systems and AI accelerator chips, designed specifically and from scratch for AI, are required.”
Note that there are many companies developing AI accelerator chips. What’s more, the focus is on narrow or weak AI. So yes, with the acquisition, Habana is looking at gaining enough scale to be a long-term winner.
“Will better performing chips be enough to change the market?” said Omri Geller, who is the co-founder and CEO of Run:AI. “While specialized hardware is a critical part of designing for an AI-first world, it’s only a part of the picture. The whole AI stack, including software like NVIDIA’s CUDA, and software used to manage and orchestrate AI accelerators, needs to be rearchitected to get the full performance out of these new chips. Ultimately, as deep learning becomes widespread and business-critical to the enterprise, it will mean that both hardware and software for AI may get rebuilt. The Habana acquisition opens the door for some interesting discussions on what to expect from the AI software stack of the future.”
Yet Intel’s move is a wake-up call—and is likely to spur more consolidation in the AI chip market. “Graphcore, Wave Computing, NUVIA, Cerebras Systems, Groq–the list could go on and is larger than you would expect,” said Leone. “Some of the companies are focused on power consumption as a key benefit, some scale, some size to fit on small devices vs. in global data centers, some are GPU-centric or FPGA-centric. The market is already ripe for consolidation and the Intel acquisition of Habana is just the start of what’s going to be an arms race to deliver specialized compute to satisfy the complexity and diversity of AI workloads.”
Tom (@ttaulli) is the author of the book, Artificial Intelligence Basics: A Non-Technical Introduction.