Intel to Acquire AI Company Habana Labs, Price Could Exceed $1 Billion

2019-12-05

According to a report on Tuesday, Intel may be in talks to acquire AI company Habana Labs for over $1 billion. This would be Intel's third acquisition of a large AI company, following Movidius and Nervana in 2016. Habana Labs is an Israeli AI company founded in 2016 by two former executives of 3D sensing company PrimeSense. PrimeSense was acquired by Apple for $360 million in 2013. Habana Labs aims to create processors from scratch for deep learning training and inference, providing higher performance, scalability, cost, and functionality. Habana Labs has raised a total of $120 million in funding and employs 120 people, with its last funding round taking place in November 2018. If this deal goes through, it would be a significant addition to Intel's product portfolio. In recent years, Intel has been investing in AI capabilities and software support from the edge to the data center, from CPUs to ASICs (Application-Specific Integrated Circuits). In 2016, Intel acquired Movidius for edge and vision AI, and Movidius recently announced the Keem Bay VPU (Vision Processing Unit). Also in 2016, Intel acquired Nervana for the deep learning training chip market for $350 million. Intel has also developed its own NNP-I for inference. Therefore, the potential acquisition of Habana Labs may indicate that there are still gaps in Intel's AI product portfolio. Habana's training chip, called Gaudi, is similar to Nervana and is based on Tensor Processor Cores (TPC) architecture, with 32GB of HBM2 and support for PCIe 4.0. It also supports popular DL formats such as INT8 and BF16. One outstanding feature of Gaudi is its interconnectivity, which can scale to hundreds of chips. It is the first AI chip with engines that can perform remote direct memory access (RDMA) through fused Ethernet, with a bidirectional throughput of 2 Tbps (via up to 10 100 Gbps Ethernet ports). RDMA means that the chip can directly access the memory of any other chip without involving the operating system. The advantages of Ethernet are that switches are easy to use and do not require PCIe switches or dedicated NICs. In terms of performance, a 140W TDP chip has a throughput of 1,650 images per second in ResNet-50, which is 14% higher than Nvidia's Tesla V100, with only half the power consumption. Based on MLPerf v0.5, a 640-chip system can achieve a throughput of 845,400,000 images per second, while a V100-based system with 80 DGX-1 can achieve a throughput of 218,300 images per second. The relative efficiency of a 512-chip Gaudi system is 80%, while a V100 system has a relative efficiency of less than 30%. Habana also claims to have its HL-1000 Goya chip, which is the first commercially available deep learning inference processor. Exact details are scarce, but it features a fully programmable VLIW TPC and supports all major frameworks. Habana claims that its performance is three times that of Nvidia's Tesla T4, with twice the efficiency and much lower latency. According to a report from Israeli media Calcalist's technology export agency CTech, Intel is in "advanced negotiations" with a potential deal value between $100 million and $2 billion. This would be Intel's largest acquisition since acquiring Israeli company Mobileye for $15.3 billion in 2017.

RECENT AI NEWS

RECENT AI TOOLS