Intel’s decision to buy startup Habana Labs in December for $ 2B set tongues to wagging about what it might mean for Nervana, the previous AI startup Intel acquired in 2016. Roughly a month later, the other shoe has dropped: Intel has announced it will cancel the NNP-T and NNP-I products it had previously planned to ship by the end of 2019 and pivot instead to focus on Habana.
There’s been little chatter about why this happened, but Karl Freund of Moor Insights and Strategy caught wind of the news. According to him, Intel will support the NNP-I for “previously committed customers,” but will cease all development on the NNP-T AI training design.
The NNP-T was designed at TSMC and intended for that company’s 16nm FinFET process and fit into 150-250W power envelopes. The NNP-I, which will apparently be supported to some small degree, is a 10-50W part that paired two Ice Lake CPU cores with 12 inference compute engines (ICE).
The thinking here is that Nervana’s hardware must not have been particularly performance competitive with what Intel’s competitors have or are bringing to market. Intel has an array of products targeting various AI, deep learning, and compute markets with Mobileye, Movidius, its upcoming Xe architecture, and its FPGA business. Habana Labs has been shipping its Goya Inference Processor since Q4 2018 and the Gaudi AI Training Processor sampled “to select customers” in the second half of 2019. Intel’s decision to essentially shut down further work on Nervana’s architecture would appear to speak for itself.
This is the second AI/machine learning effort Intel has shut down, after Xeon Phi, but I’m not sure how much I’d read into that. Xeon Phi was an effort to create a new x86-based product that could compete with GPUs in double-precision workloads. Intel’s 10nm issues prevented the company from pushing Xeon Phi to lower nodes, but the roots of the architecture date back to Intel’s failed Larrabee GPU experiment, not any kind of determined effort to build an AI/ML processor.
Canceling Nervana’s products undoubtedly hit those product teams hard, but we’re still very early in the AI game, and a lot of companies are currently working on first-gen accelerators. The fact that Goya was in-market and Gaudi was sampling before Intel bought the company provides some assurance that the CPU manufacturer won’t have to spend a huge amount of time bringing an initial part to market.
AI and machine learning have been pretty Nvidia and Intel-centric up until now, and that trend seems likely to continue for the near-term future. AMD’s presence in these markets — and its ability to compete effectively via translation projects like ROCm, compared to native support for Nvidia CUDA — have both been modest in comparison to its largest rivals.
To be fair to AMD, that’s not an accident. When analysts have asked whether AMD would be explicitly targeting the emerging AI/ML space, AMD’s executives have generally said that they would do so in a limited way and with specific products when it made sense to do so. AMD isn’t making any effort to target CPUs like the Epyc 7742 for AI workloads the way Intel focused on adding AVX-512 support to Xeon.
2020 will be a significant year for the AI/ML market. The first consumer iteration of Intel’s GPU will hit markets this year. While these will obviously be intended for mainstream systems, you can bet analysts will have an eye towards any data center-friendly features or capabilities that might make an early appearance. Nvidia’s next-generation 7nm GPUs are expected to drop this year, as is AMD’s Navi 20. It’s not crazy to think Navi 20 could tip up in new server, workstation, or data center GPUs from AMD as well this year.
In short, pivoting towards Habana and away from Nervana doesn’t necessarily mean Intel is falling behind — not so long as Habana’s performance and roadmap are better aligned with Intel’s customers than what it was fielding previously.