Deep learning has been a hot topic this year, with high-profile announcements from companies like IBM, Facebook, Google, Nvidia, Qualcomm, and Tesla. Now, Intel is tossing its own hat into the ring by buying the deep learning software and hardware developer Nervana Systems.
Nervana Systems has a cloud-based AI system that it sells to customers who want to adapt deep learning for their own specific use-cases and businesses as well as a proprietary, GPU-specific framework dubbed Neon. The company’s third product hasn’t actually launched yet, but it might be the principle reason why Intel bought this company in particular. The Nervana Engine is an ASIC that focuses on the strengths of what GPUs bring to the table, rather than the not-insignificant amounts of hardware that ultimately aren’t useful to deep learning problems.
The reason that GPUs are useful for these kinds of applications is because they contain enormous arrays of cores that can be employed to solve complex problems. Resources like ROPs, texture caches, and FP64 (or even FP32) support aren’t particularly important for deep learning, however — that’s why Pascal’s 16-bit half-precision mode was something Nvidia talked up when it unveiled GP100 earlier this year. Nervana’s existing Neon engine already runs on Nvidia hardware, but Intel’s decision to buy the company will likely put an end to more permissive licensing arrangements.
Right now, Intel is stuck in a bit of a tight spot. The company’s consumer revenues have declined alongside the PC market’s downturn, but its data center and HPC markets remain quite healthy. Intel missed out on the entire mobile and tablet market, and already had to cancel its plans to create new business for itself in those spaces (a failure we chronicled in a two-part article earlier this year).
This goes beyond not wanting to miss an emerging market, however. Intel has been acquiring companies with product lines and markets that stretch beyond its own dominance of the data center, consumer, and high performance computing markets. While products like Xeon Phi could theoretically be used for deep learning, Xeon Phi is designed to perform massive vector calculations, not the half-precision operations that a deep learning network uses. It also packs far fewer cores than an Nvidia Tesla or even an equivalent AMD card, though we’d caution against treating core counts as indicative of deep learning performance.
If deep learning is as central to the future of AI and computing as the industry has claimed, entering the market by acquiring a company with specialized ASIC hardware and proven designs is an excellent way for Intel to ensure that it remains relevant as computing continues to evolve. It could also be read as a tacit admission that Intel isn’t necessarily sure how to continue to push the evolution of microprocessors farther than it has already.
I’ve talked before about how Intel isn’t just dragging its feet on Moore’s law — there are fundamental limits to silicon engineering, and they aren’t going away. Moves like this could be read to mean that even Intel recognizes that the era of huge advances in general purpose compute performance are mostly over. Machines will continue to draw less power and be slightly more efficient over time, but the last major leap for Intel’s CPUs was Sandy Bridge over Nehalem. Haswell and Skylake were much more modest improvements.
Moving into markets like this gives Intel the opportunity to explore other types of compute architectures, not as replacements for x86, but as high-performance supplements to it.