You are currently viewing SemiWiki as a guest which gives you limited access to the site. To view blog comments and experience other SemiWiki features you must be a registered member. Registration is fast, simple, and absolutely free so please, join our community today!

Results 1 to 4 of 4

Thread: Huawei to Have NPU in Phone Chip

  1. #1
    Join Date
    Apr 2013
    East SF Bay Area
    Thumbs Up
    Received: 403
    Given: 345

    Huawei to Have NPU in Phone Chip

    Huawei is going to introduce an neural processing unit in their new mobile chip and it looks like the arms race is on. It won't be true AI for it unloads some of the work to the cloud. They have made recent claims in the last few days that this chip out performs the new Apple chip. Only a side by side comparison will tell. It's built on a 10nm node, but no word on what foundry builds it.

    AI and the cloud crunch the numbers for Huawei's new mobile chip

    0 Not allowed!
    Last edited by Arthur Hanson; 09-05-2017 at 06:50 AM.

  2. #2
    Join Date
    Feb 2016
    Thumbs Up
    Received: 51
    Given: 46
    Foundry is TSMC. Density is 55 Mtr/mm2. Transistor count is 5.5 bilions.

    0 Not allowed!

  3. #3
    Admin Daniel Nenni's Avatar
    Join Date
    Aug 2010
    Silicon Valley
    Thumbs Up
    Received: 860
    Given: 2,309
    Huawei's chip unit HiSilicon is a big TSMC customer. They have a decent sized design group in Hsinchu. It is kind of funny because there are no signs for it, very stealthy but we all know it is there....... HiSilicon licenses the ARM architecture like Apple and QCOM so they are definitely a force to be reckoned with. According to IDC:

    Huawei sustained its dominance in China growing nearly 23% as shipments climbed from 28.1 million units last year, to 34.6 million units in the first quarter of 2017. Huawei once again demonstrated its stable position in the premium market with the P and Mate Series, and a strong presence in the affordable sector with its Y Series and Honor brand. Although Huawei announced earlier in the month that the Mate 9 has sold over 5 million units since it launched in November, in the U.S. the device, as well as the brand, has failed to grab consumer attention. This U.S. attention is something they will need if they aspire to displace the two market leaders. The launch of the new P10 flagship and the new P10 Plus at the very end of the quarter presents consumers a valid third option (outside of Apple and Samsung) for the coming quarter thanks to both premium design and similar performance.

    IDC: Smartphone Vendor Market Share

    0 Not allowed!
    Now available in print or Kindle: "Mobile Unleashed: The Origin and Evolution of ARM Processors In Our Devices"

  4. #4
    Top Influencer
    Join Date
    Sep 2016
    Thumbs Up
    Received: 77
    Given: 21
    "They have made recent claims in the last few days that this chip out performs the new Apple chip."

    No they didn't. What they said was that for image recognition Kirin 970 runs at just over 4x the speed of iPhone 7 Plus. That's the existing (one year old) iPhone.
    How could they possibly have performance data against the A11?

    Apple did everything but announce that they are including, let's call it neural hardware, in the A11 --- lots of references to machine learning APIs and how using them would make optimal use of current and future hardware.
    The only real question with the A11 is whether there is specifically NPU hardware that's good for nothing else, or whether the SoC comes with a somewhat generic "throughput engine" that handles everything from GPU to neural network to generic multi-core processing.

    It's certainly true that Kirin 970 has a lot of transistors on the chip, 5.5 billion, though they gave us no die shots or indication of where the transistors go (eg how large are the NPU and GPU). So they're probably still going to look reasonable compared to Apple, but my guess is that A11 will exceed them in this particular task (2000 images per minute).

    [One issue that isn't clear is the extent to which like is being compared with like. If Apple are targeting say 98% accuracy and Huawei 95%, or Apple are targeting 4000 possible identification tags and Huawei 2000, that may have significant implications for performance. We don't yet really have any good benchmarks for this space, or even know what we should be looking for.]

    But certainly it's good to have some competition keeping Apple aggressive. (cf Intel and what happens in the absence of competition.)

    0 Not allowed!

Tags for this Thread

Posting Permissions

  • You may not post new threads
  • You may not post replies
  • You may not post attachments
  • You may not edit your posts