Intel is planning another big acquisition.

According to calcalistech Report a>, a person familiar with the matter revealed that Intel may be negotiating the acquisition of Israeli AI startup Habana Labs with a valuation between 1 billion and 2 billion US dollars. This will be Intel’s third acquisition of a large AI company after Movidius and Nervana in 2016.

This is the second time Intel has acquired a technology company in Israel. In 2017, Intel acquired the autonomous driving company Mobileye for $ 15.3 billion, the largest acquisition in Israeli technology history.

Where is this company sacred?

Habana Labs was founded in 2016 and is headquartered in Tel Aviv, Israel. According to official information, currently The company has 150 employees worldwide and has offices in Tel Aviv, San Jose, California, Beijing, and Gdansk, Poland.

The core founding teams are all “old guns” in the technology industry. The CEO and co-founder is David Dahan, and the vice president of R & D is Ran Halutz. The two founders’ previous entrepreneurial projects. The 3D sensing company PrimeSense was approved by Apple in 2013. Acquired for $ 360 million.

Avigdor Willenz, chairman of Habana Labs and early major investor, was the founder of chip provider Galileo, a company Acquired by Marvell for $ 2.7 billion; Annapurna Labs, which was subsequently established, was acquired by Amazon for $ 370 million.

In short, this team is a team with abundant technology, capital and experience.

Legend: Avigdor Willenz

Habana’s main business is the design of chip wafers, which is equivalent to the “midstream” of the chip industry. The main direction of attack is to customize AI chips for deep neural network training and inference.

AI chips can be specifically divided into training chips for building neural network models, and inference chips that use neural network models for inference and prediction. In 2019, Habana released two AI chips-the inference chip Goya and the training chip Gaudi.

According to the official website, Goya’s performance is three times that of Nvidia’s Tesla T4 and its efficiency is two times. Facebook used Goya to compile the Glow ML compiler.

The new training processor solution introduced in June this year is called Gaudi (PDF), and its architecture is based on the Tensor processor core with a two-way throughput of 2Tbps.

MLPerf (an international benchmark for measuring and improving the performance of software and hardware for machine learning), jointly released by many global technology giants and universities, announced this year’s “inference chip benchmark scores based on data submitted by 14 institutions “.

Among them, in addition to Nvidia’s product category in a number of commercial equipment scenarios, Habana Labs, Google and Intel also have the highest scores.

Especially in the field of data centers, Nvidia’s closest competitor is Habana Labs’ inference chip Goya.

Chip analyst Karl Freund said: “Habana is the only challenger capable of fully producing high-performance processors. When the next-generation MLPerf benchmark dimension is expected to include power consumption data, Habana should perform well.”

In an interview with the authoritative electronic technology website EETimes, Habana pointed out that the benchmark score is currently purely based on performance-power consumption is not a measure, and practicality (such as considering whether the solution is passively cooled or water cooled) is not, cost Of course not.

It is worth mentioning that Habana Labs just completed $ 75 million in financing last November, when Intel was the lead investor. The investment was not fun, and the whole acquisition came simply.

Why buy?

Has previously acquired two AI startups.

In August 2016, Intel acquired AI startup Nervana and announced its entry into the field of deep learning training chips. After more than three years of preparations, in late November this year, Intel released the first cloud AI custom chip-Nervana Neural Network Processor (NNP) series.

As mentioned above, custom chips are geared towards specific user needs, so they have stronger performance and lower power consumption in dedicated areas. They are widely used in cloud server deployments and are considered to be one of the most promising areas. One.

Intel and NVIDIA have been leaders in this field. Especially in the server field, 98% of the world ’s data centers use Intel processors, and downstream cloud service providers such as Amazon and Microsoft are customers of Intel chips. As a result, AI chips have become a seller’s market with prices Stay high.

In order to break this “monopoly,” Amazon and Google have begun to enter the ASIC custom chip. Amazon, the largest player in cloud services, launched the first AI chip, Inferentia, as early as last year. During the same period as Intel’s release this year, Amazon’s chips have iterated to the second generation, which is at least 20% faster than the first generation.

Google is no different. Although a different solution is used, Google’s Tensor Processing Unit (TPU) has iterated to the third generation.

Cloud computing is still an industry that burns money. According to the latest IDC data, global spending on artificial intelligence systems in 2019 is expected to be $ 37.5 billion, and by 2023, this number will reach $ 97.9 billion.

Habana labs is a fairly high-volume acquisition, and this is Intel’s third acquisition of an AI startup. In fact, it is the consistent strategy of American technology giants to deepen the layout and make up for shortcomings through acquisitions. M & A in the US technology industry is quite active.

From the current AI custom chip market, Intel has reason to feel pressure.

Investment Bank Bernstein analyst Stacy Rasgon once pointed out that Amazon’s first ARM chip does not seem to have an impact on Intel’s data center business, which has continued to grow in the past year .If Amazon or Google invest in ARM technology and master these technologies, this may cause more trouble for Intel.

As a result, Intel needs to continue to strengthen its technology to prevent former downstream customers from “single flight” or even becoming their competitors. On the consumer side, AMD has become a threat that Intel cannot ignore. On the B side, Intel wants to stifle competitors.