Intel is making a huge push into AI and deep learning, and intends to build custom variants of its Xeon Phi hardware to compete in these markets. Several months ago, the Santa Clara corporation bought Nervana, an AI startup, and this new announcement is seen as building on that momentum. AI and deep learning have become huge focuses of major companies in the past few years — Nvidia, Google, Microsoft, and a number of smaller firms are all jockeying for position, chasing breakthroughs, and building their own custom silicon solutions.
The upcoming Knights Mill is still pretty hazy, but Intel has stated that the chip will be up to 4x faster than existing Knights Landing hardware. Right now, the company is working on three separate forays into the AI / deep learning market. First up, there’s Lake Crest. This product is based on Nervana technology that existed prior to the Intel purchase. Nervana was working on an HBM-equipped chip with up to 32GB of memory, and that’s the product Intel is talking about rolling out to the wider market in the first half of 2017. Lake Crest will be followed by Knights Crest, a chip that takes Nervana’s technology and implements it side-by-side along with Intel Xeon processors.
“The technology innovations from Nervana will be optimized specifically for neural networks to deliver the highest performance for deep learning, as well as unprecedented compute density with high-bandwidth interconnect for seamless model parallelism,” Intel CEO Brian Krzanich wrote in a recent blog post. “We expect Nervana’s technologies to produce a breakthrough 100-fold increase in performance in the next three years to train complex neural networks, enabling data scientists to solve their biggest AI challenges faster.”
To date, the companies that have done well with AI — well, company — has been Nvidia, whose GPU technology is powering a great deal of cutting-edge R&D. Claims that Intel needs a specific GPU architecture to compete, however, are mistaken. GPUs are good at these kinds of computing projects because the projects map well on to the hardware we use for gaming — not because there’s something magic about graphics processors that makes them uniquely and specifically suited to the tasks. Put differently, you could build a GPU-style compute engine without any of the IP blocks or hardware that transform it into a graphics card.
Xeon Phi began life as a GPU (albeit a GPU with a very different focus than cards from AMD or Nvidia) and was reinvented into a vector processor. There’s nothing to say Intel can’t bend it back a bit, possibly by building lower-precision registers or offering them as options on certain types of hardware. Deep learning and AI typically use much less precision than other types of workloads; Intel CPUs support the IEEE 754 floating point standard and can offer up to 80 bits of precision, while most deep learning and AI workloads are done with 8-bit or 16-bit calculations.
AMD is also dipping a toe into this business area via GCN, but we don’t know yet if deep learning and AI will have an impact on the company’s upcoming Vega architecture. Most of AMD’s focus remains on the gaming market, where its console wins have been critical to shoring up the company’s business.