Intel pushed into the world of artificial intelligence: Habana Labs acquired for 2 billion

Intel has announced the acquisition of Habana Labs, an Israeli company engaged in the development of accelerators programmable for deep learning for datacenters. The economic outlay is approximately $ 2 billion.

“This union strengthens the artificial intelligence portfolio and accelerates efforts in the nascent – but strongly developing – market dedicated to artificial intelligence chips, “read an Intel note. The Santa Clara house believes that this sector will reach a turnover of over 25 billion by 2024.

Navin Shenoy, vice president and general manager of Intel's Data Platforms Group, explained that this is a further step in the Intel's strategy in the field of artificial intelligence, devoted to providing customers with solutions for every performance need.

“Habana turbocharges our offer of AI solutions for datacenters with a family of processors for training high-performance neural networks and a standards-based programming environment that deals with ever-changing artificial intelligence loads.”

Intel expects to bill over 3.5 billion in the AI ​​market for the current year, with growth of over 20% on an annual basis. “We know that customers are looking for easy programmability with specially designed AI solutions, as well as superior and scalable performance on a wide variety of workloads and neural network topologies. That's why we are excited to integrate a team of Habana's caliber with a proven track record. The combined skills and intellectual properties will offer unparalleled processing performance and efficiency for AI workloads in datacenters. “

Habana will remain an independent unit and continue to be led by current leaders. Habana will report to Intel's Data Platforms Group. President Avigdor Willenz has accepted the position of senior adviser to the division and Intel. The company will continue to operate in Israel, where Intel has a strong presence with research centers and factories. Prior to the acquisition, Intel Capital, the US company's financial arm, invested in Habana.

The company developed the Gaudi processor for training neural networks, currently being tested by some customers in the hyperscale sector. According to Intel, Gaudi-based systems should guarantee up to four times the throughput increase compared to systems “based on an equivalent number of GPUs”.

Gaudi can process 1650 images per second with a batch size of 64 by training a ResNet-50 network. Performance that has been achieved with a consumption of 140 watts. Both chips have eight very long instruction word, single instruction multiple data (VLIW SIMD) vector cores, which Habana calls the tensor processor core (TPC).

One of the interesting aspects of Gaudi is the RoCE (Remote direct memory access over Converged Ethernet) network integrated in the chip. In practice, it offers 10 Gigabit Ethernet 100 ports directly on the processor, which facilitates scalability and allows very complex training systems with dozens of chips based on non-proprietary standards-based interfaces.

Habana has also developed a solution for inference, called Goya, already available. Goya was presented in September 2018 and can process 15,000 images per second via the ResNet-50 neural network with a batch size of 10 (more than 5 times the number of images of competing platforms) with a latency of 1.3 ms consuming 100 watts.