Moor Insights & Strategy analyst Karl Freund told Forbes that Chipzilla had decided to end its work on both the Nervana NNP-T training chips and the Nervana NNP-I inference chips, though it said it will still deliver on customer commitments.
Habana has developed two AI chips of its own, namely the Habana Gaudi and the Habana Goya. Gaudi is a highly specialised neural network training chip, while Goya is a processor used for uses neural networks in active deployments.
Chipzilla thought that both were built for more or less the same purpose, so it doesn’t make sense for the company to build out both product lines.
Analysts had pondered the future of the Nervana chips at the time Intel bought Habana. Freund said it was “hard to imagine” a scenario where Nervana processors would continue to play a significant role in Intel’s portfolio.
Habana chips are more powerful. In recent benchmark tests, two Nervana NNP-I chips racked up 10.567 inputs per second in ResNet-50. But just one Habana Goya chip was able to reach 14,451 inputs per second in the same test. Goya is not only more powerful, but it has been shipping to customers since 2018, while NNP-I is yet to be released.
Intel bought Nervana Systems for around $350 million in August 2016 in order to crack the market for deep learning training chips, which was up until then dominated by Nvidia with its more powerful graphics processing units. The thinking at the time was that by developing specialised application-specific integrated circuits for deep learning, it could gain a competitive advantage over Nvidia.