In a 2 Billion Dollar Deal, Intel Acquires AI Chipmaker Habana Labs
In a bid to galvanize its own AI future, Intel has acquired advanced AI training and inference silicon technology.
Intel has acquired deep-learning powerhouse Habana Labs in a bid to galvanize its own AI future. This announcement accompanies a push for expansion from Intel, which has long had a prominent presence in data centers. Intel believes this $2 billion investment will help it penetrate a projected $25 billion AI chipset market by 2024.
Habana Labs’ Gaudi (left) and Goya (right) units. Image from Habana Labs
Habana Labs claims that its solutions are modular, ethernet-powered, and easy to rack mount.
The Appeal of Habana Labs
What makes the Israeli hardware vendor such an attractive business asset? The company’s groundbreaking AI training and inference silicon—dubbed Gaudi and Goya, respectively—are said to markedly excel at their intended tasks.
Gaudi HLS-1. Screen capture used courtesy of Habana Labs
Gaudi and Goya adhere to connectivity standards. Habana Labs also states that these scalable technologies play nicely with existing infrastructure without the need for reconfiguring. In addition, the company asserts that these units are future-proofed and primed to tackle evolving workloads.
"Gaudi," AI Training Processor
Because AI models improve through continued exposure to data sets and graphics, even simple errors can accumulate, leading to a costly and time-consuming correction process.
Habana Labs seeks to prevent this issue with its deep-learning training solution, (PDF) Gaudi.
The Gaudi HL-2000 chip, included in the Gaudi platform, is available as a PCle card or a Mezzanine card; these two options allows for compatibility in small-scale servers and medium- to large-range systems alike.
Gaudi HL-205 Mezzanine Card (left) and Gaudi HL-200 PCIe Card. Image used courtesy of Habana Labs
The chip features a processor die and four HBM memories.
Each Gaudi chip achieves peak performance while only drawing 140 watts of power. This energy reduction may account for noticeable cost savings.
Gaudi compared to Nvidia V100. Image (modified) used courtesy of Habana Labs
Habana claims that Gaudi outpaces Nvidia V100's throughput by 3.8 times when comparing groupings of ~650 processors.
"Goya," Deep Learning Inference Platform
Habana also created (PDF) Goya, an inference engine that includes a programmable Tensor Processing Core, development tools, libraries, and a compiler. According to Habana Labs, Goya is capable of "massive data crunching with low latency and high accuracy."
As with other inference platforms, Goya can interpret facts and apply new rules to incoming data. However, Goya accomplishes this at an impressive rate—achieving 1,527 sentences per second of throughput on BERT.
A single Goya PCIe can classify 15,453 images per second on RESNET-50 with almost no latency. That automated processing is said to slash through the noise associated with big data.
RESNET-50 throughput and latency. Image (modified) used courtesy of Habana Labs
Combatting Stagnation and Competition
Intel’s acquisition is aptly timed in response to competing developments earlier this year. AMD’s Epyc processors have risen in popularity thanks to AMD’s latest deal with Alphabet. Major cloud companies like Amazon and Microsoft are also adopting this technology within their data centers.
Intel must present compelling hardware now more than ever to preserve its industry standing. Habana Labs may be the key to offering the AI software development tools to enhance Intel's hardware compatibility; this move, specifically, extends the performance of its CPU-GPU pairings.
Intel’s History with AI
Intel has long been a CPU provider for demanding systems—systems that thrive when hardware compatibility is fully optimized.
Some of these same systems pair Intel’s Xeon CPUs with Nvidia’s Tesla P100 and V100 GPUs – which Habana Labs claims to have trounced at scale, according to benchmarks.
Modern data centers power today’s cloud applications. They also store massive amounts of qualitative and quantitative data locally. Intel provides the hardware behind the vast majority of these services.
Avigdor Willenz, chairman of Habana Labs, with a rack that uses Habana Labs’ HLS-1 Gaudi AI training system. Image used courtesy of Intel
As Intel continually introduces AI to cloud applications, Habana Labs AI-forward processors may be a helpful support tool.
Where Does Intel Go from Here?
Since Intel has access to Habana Labs' resources, we may expect some architectural changes to support Gaudi and Goya. Intel claims their relationship with Habana will be mutualistic. Intel will have access to their new team’s hardware, and Habana will have access to Intel’s extensive AI library.
This will hasten development for both parties, ensuring software and hardware will simultaneously grow more sophisticated. Developers and engineers will collaborate instead of working in siloed teams, which in turn will promote product harmony.