Datasets:

---

Recag

/

Rp_CommonC_12

like 0

ERROR: type should be string, got "https://www.barrons.com/articles/intels-a-i-director-singer-lays-out-the-vision-for-deep-learning-1525901912\nTechnology Trader\nIntel’s A.I. Architect Singer Lays Out the Vision for Deep Learning\nTiernan Ray\nUpdated May 9, 2018 5:38 pm ET / Original May 9, 2018 1:58 pm ET\nIntel A.I. exec Gadi Singer.\nOn Tuesday, I swung by Intel’s (INTC) campus in Santa Clara for a most interesting discussion with GadiSinger, an Intel vet and the person leading the company’s design of computer chip \"architecture\" for artificial intelligence.\nSinger has been through numerous chip innovations with Intel, including running the team for the original “Pentium” processor, technically the “586,” and server chips such as the “Itanium” and the first “Xeon” processor.\nSinger laid out for me the thinking at Intel about A.I. in advance of the company’s first-ever dedicated A.I. conference, which happens next month.\nWith A.I., Intel faces a challenge: the dominant player by conventional wisdom is Nvidia (NVDA), whose graphics chips, or “GPUs,” have been widely used by Alphabet’s (GOOGL) Google and Facebook (FB) and all the other cloud companies to do much of the heavy lifting for machine learning and deep learning. In addition, Silicon Valley has cultivating numerous startups focused on dedicated A.I. chips. Advanced Micro Devices (AMD), Intel’s rival in PC chips, is also pushing the use of GPUs for A.I., being the only other vendor of size in GPUs besides Nvidia.\nGoing up against that, Intel has Xeon, and also field-programmable gate arrays, acquired with the purchase of Altera, and also more specialized A.I. chips of its own from the acquisition a couple years back of Nervana. (Singer reports to Nervana co-founder Naveen Rao, who heads up Intel's A.I. efforts.)\nSinger offered two points in particular about the A.I. market that seem to encapsulate how Intel is thinking about the competitive situation.\nOne is that more and more, A.I. scientists are focusing on the “frameworks,” sets of software that structure one’s approach to machine learning and other A.I. approaches, with names such as “TensorFlow\" and “MXnet” and \"Caffe.\"\n\"The cloud service providers prefer the model of programming directly to deep learning frameworks,” contends Singer. “And academia is more and more innovating on top of those frameworks.” What that means is that Nvidia’s lead in software, with its “CUDA” technology, developed over a decade at the company, is not the edge that many people ascribe to it. People don’t want CUDA as much as they want these open-source, widely used frameworks, he contends.\nIntel flag flying outside its Santa Clara headquarters.\nTo address that, Singer says that what Intel is doing is to provide more and more software that improves the applicability of Xeon for many A.I. tasks:\nThe thing that was completely missing [in Xeon] is an optimized software stack. When the first discussions around deep learning started in the industry in 2016, the Xeon libraries were not optimized for deep learning. We have seen some 100-times improvements since then once we mapped the basic primitives of deep learning into the instruction-set architecture of the Xeon, of the FPGAs, and also the Nervana chips, and those from Movidius. We are not there yet, but we are on track to make that better and better.\nThis software lesson is one of the things that was impressed upon Singer from the experience with Itanium. Although Itanium shipped in systems and had enthusiasm from some partners such as Hewlett-Packard, the view from many chip-industry observers over the years was that it was a failure. Singer concedes that there was not enough work done on software to make the Itanium all that it could have been. His focus on software with A.I. seems to be an effort to avoid that fate this time around.\nThe second point that Singer stressed again and again is that for A.I., there will be a variety of tasks, some of which happen in the traditional “sequential” manner of computing — the chip loads one instruction after another -- some of which are “spatial” in nature, meaning that they go back and forth between inputs and outputs to an algorithm.\nSinger believes the breadth of Intel’s chips will better enable researchers to perform tasks across that spectrum of kinds of compute. By contrast, Nvidia’s GPUs are less than ideal for most A.I., he contends:\nGPUs are not the best structure to do deep learning. They're pretty good, but they are not optimized. Think about the underlying concept of temporal versus spatial architectures. In a temporal architecture, you have a flow of instructions, taking data from a known place to a known place. With neural networks, there is a wave front that flows through a very wide set of operations, and you go through all those multiple nodes. Architectures that are more spatial have an advantage in that they don't have to rely on a flow, and they don’t end up waiting on other instructions. The GPU is not the optimal architecture for that kind of compute.\nSinger goes on that in the “inference” stage of computing, when A.I. questions are answered by the computer, as opposed to the “training” phase, when the computer is first learning the rules of a problem, \"we see blended compute types.”\n\"Applications such as machine translation, or mapping, like a genome sequence, are a combination of neural nets and things that are more traditional branching operations,” again, a more-traditional kind of compute married to instructions that are particular to deep learning.\nAgain he emphasizes, “the GPU is not the strongest in doing this type of branching,” meaning, the more-traditional part at which Xeon excels.\nIn a sense, Singer was making the statement that it is Intel's design of its chips, not just their manufacturing quality, that will win the battle in A.I. That is a significant contention given that the chip industry has for years presumed Intel's design of its chips has not kept pace with the needs of newer forms of computing such as A.I.\nIntel shares ended today up 71 cents, or 1.3%, at $54.34.\nSign up to Review & Preview, a new daily email from Barron’s. Every evening we’ll review the news that moved markets during the day and look ahead to what it means for your portfolio in the morning.\nOn Tuesday, I swung by Intel’s (INTC) campus in Santa Clara for a most interesting discussion with GadiSinger, an Intel vet and the person leading the company’s design of computer chip "architecture" for artificial intelligence."