Slashdot Mirror


Huawei Unveils AI Mobile Chipset Said To Rival A11 Processor In Upcoming iPhones (macrumors.com)

On Saturday, Chinese mobile maker Huawei unveiled its first artificial intelligence smartphone chipset, which it hopes will lure customers away from Apple's upcoming range of new iPhones and towards the Asian company's "most powerful handset yet," the Mate 10, which is set to debut next month. Mac Rumors reports: Huawei touted the Kirin 970 AI mobile chipset's built-in "neural processing unit" at the IFA consumer electronics trade show in Berlin, claiming that the technology is "20 times faster" than a traditional processor. The world's third largest smartphone maker claimed that mobile devices powered by the Kirin 970 will be able to "truly know and understand their users," by supporting real-time image recognition, voice interaction, and intelligent photography with ease. According to Nikkei, the Kirin 970 integrates 5.5 billion transistors in a single square centimeter about the size of a thumbnail, which includes an octa-core central processing unit, a 12-core graphics processing unit, a dual-image signal processor, a high-speed 1.2Gbps Cat.18 modem, and AI mobile computing architecture. The Kirin 970 is said to be based on the same 10-nanometer technology as Apple's existing A10X Fusion processor and the A11 processor that will power its new iPhone range, set to debut this month. The Mate 10 is said to be a bezel-less all-screen handset with a 6-inch, 2:1 display and a 2,160 x 1,080 resolution. Like Apple's so-called "iPhone 8," the Mate 10 is also expected to feature some form of facial recognition and improved cameras.

1 of 77 comments (clear)

  1. Re:NPU - The First Neuron by ShanghaiBill · · Score: 3, Informative

    I'm not sure how NPUs are implemented, my guess is you have an array of gates like a FPGA

    Most likely it is a FinFET ASIC with local memory and lots of FP16, FP12, and/or FP8 multipliers for doing matrix ops. That is what Google's TPU is, and this is targeted to the same apps (deep NNs).