In December 2019, Intel had announced its decision to acquire a startup — Habana Labs for a whopping $2 billion. The acquisition sent a signal what it might mean for Nervana — another AI startup acquired by Intel in 2016. Now, Intel is cancelling the NNP-T and NNP-I products that were launched in November 2019.
For Intel, the focus will now be on Habana’s HL-x, a series of neural processors for both training and inference in the data centre. There are two separate chips based on two different microarchitectures, one for inference and one for training — similar to what had been the case with Nervana. This includes Goya, a 16-nanometer microarchitecture for inference neural processors, and Gaudi, which was designed as a microarchitecture for the acceleration of training in the data centre. Both HL series are intended to have a common unified architecture, which makes software development simple — another thing which makes it similar to Nervana neural network processors.
It is to be noted that Intel acquired Nervana in 2016 for $400 million, and then appointed Nervana’s CEO Naveen Rao to lead its artificial intelligence products group. The idea was to develop a specialised application with specific integrated circuits for deep learning, to gain a competitive advantage over Nvidia.
Reason Behind Intel Cancelling Nervana Products
There have been discussions uncovering the possible reasons on why this took place. Karl Freund of Moor Insights and Strategy first broke the news and gave the clear-cut reason for the cancellation of Nervana processors in a blog. According to Freund, Intel will continue to support NNP-I for earlier committed customers, and it will stop all other development related to NNP-T chip design.
According to experts, Intel’s Nervana chips were not at par with Habana in terms of performance and redialing their efforts to Habana chips helps the company to take on other prominent players in the AI hardware market such as Nvidia, Samsung and AMD.
Habana Labs has shown its Gaudi AI processor, which the company claimed can outperform a GPU by a factor of 4x. Habana also exhibited great results in the latest round of MLPerf inference benchmark scores where it was only second to Nvidia in some categories, something that Intel would have considered thoroughly as it scrapped Nervana for Habana chips.
According to experts, what makes Habana chips deliver better performance is its on-chip RoCE (Remote direct memory access over Converged Ethernet) network that helps rapid scalability that is needed to build and train a neural network.
What This Means For Intel & AI Accelerator Market
According to analysts, the move towards Habana is a positive step for Intel as long as Habana’s performance and roadmap are better aligned with Intel’s customers.
Nervana neural network processors were just one of Intel’s wide range of products for AI, machine learning and deep learning market segment. In addition to its multi-purpose CPUs and specialised CPUs like Xeon processors with DL boost, the company owns assets like Movidius, Altera, Mobileye.
Also, Intel’s upcoming Xe architecture, along with its FPGA business, is set to disrupt the AI accelerator market across the globe. Here, Habana would play a critical role in strengthening the hardware. Overall, 2020 will be an important year for the AI and ML hardware market, with launches such as Nvidia’s next-generation 7nm GPUs, AMD’s Navi 20 data centre GPUs, and the first consumer version of Intel’s GPU.
When it comes to the competition between Nvidia and Intel, both companies have made acquisitions that will add to their total addressable markets (TAM) — Intel’s $ 2 billion Habana purchase and Nvidia’s Mellanox acquisition for $6.9 billion.
“It made no sense whatsoever for Intel to acquire Habana and yet hang on to Nervana. It already owns Xeon, Altera, MobileEye, Movidius and coming GPUs. Something had to give, and that was Nervana,” wrote Karl Freund.