08:30AM EDT – Real time computer vision, low power AR, language translation
08:30AM EDT – (so the next smartphone will have a big new camera feature I imagine)
08:30AM EDT – Use AI for: Task Scheduling, load balancing, memory allocation, real-time painting, ui rendering, graphics processing, camera image processing
08:29AM EDT – ‘Open architecture’
08:29AM EDT – > rather than just Huawei’s apps
08:29AM EDT – If it uses an Android standard, that might make it used by other apps
08:28AM EDT – Uses Kirin AI API and Android NN
08:28AM EDT – Open AI Ecosystem: support for Tensorflow / Tensorflow Lite, Caffe/Caffe2
08:28AM EDT – CPU took 120 seconds, so 20x perf
08:27AM EDT – Video demo showing 200 image recognition done in 6s on NPU
08:27AM EDT – That’s only using the S8 CPUs
08:27AM EDT – 2005 images per minute using CPU+GPU+TPU, comapred to Samsung S8 CPU which does 95 per minute
08:26AM EDT – 0.3W for image recognition. 1000 images took 0.19% of a 4000 mAh battery
08:26AM EDT – (So NPU is the size of four CPU cores?)
08:25AM EDT – High performance density = 25x performance in half the CPU size
08:25AM EDT – NPU can do image processing in realtime
08:25AM EDT – ‘All benchmark data is based on internal testing’
08:25AM EDT – Perf of NPU = 25x CPU. Efficiency of NPU = 50x CPU
08:24AM EDT – no wait, that’s just 10nm
08:24AM EDT – 20% power consumption reduction over Kirin 960
08:23AM EDT – ‘4.5G’ again
08:23AM EDT – Cat 18 LTE, 1.2 Gbps
08:23AM EDT – Using TSMC10nm – 10 years ago was using 180nm
08:22AM EDT – NPU follows the ISP, DSP, coprocessor as the next element needed in smartphone chips
08:22AM EDT – First company to have a dedicated AI processor inside
08:22AM EDT – ‘Dual 4.5G LTE’ technology ?
08:22AM EDT – Dual ISP
08:22AM EDT – ‘4.5G’ LTE Modem
08:21AM EDT – 5.5 billion transistors
08:21AM EDT – Image DSP with 512-bit SIMD, 32bit/384k HiFi Audio, i7 sensor processoer
08:20AM EDT – Kirin 970
08:20AM EDT – ‘Our Solution’
08:20AM EDT – Huawei wants to invest in high performance on-device AI for the best intelligent experience
08:19AM EDT – When a smartphone is detecting an object, it needs a lot of power
08:19AM EDT – We need a TPU in the a power consumption for a smartphone and a limited battery. 16 GFLOPs at 0.4 mA
08:18AM EDT – Needs high performance and high efficiency
08:18AM EDT – Stage 4: Power
08:17AM EDT – AI can assist with on-device security
08:17AM EDT – On device processing gives more privacy protection (when done right)
08:17AM EDT – 3: Security
08:17AM EDT – Recognising the scene and the behaviour
08:16AM EDT – This can be complex for a smartphone beyond simple data entry
08:16AM EDT – Stage 2: Cognition
08:16AM EDT – Can a smartphone detect a Human Being’s emotion?
08:16AM EDT – e.g. one microphone going to two, to an array with audio processing as one input to the AI. Need to identify the important information
08:15AM EDT – Having more angles to learn is important for AI, and you need the processing power
08:15AM EDT – Perception on a smartphone is the camera and the microphone
08:15AM EDT – 4 challenges: Perception, Cognition, Security and Power
08:14AM EDT – Today focusing more about on-device AI
08:14AM EDT – Individual Intelligence from the device, and collective intelligence from the cloud
08:14AM EDT – Individual Intelligence from the device, and collective intelligence from the cloud
08:12AM EDT – >He just quoted F1. Do Huawei sponsor someone in F1?
08:12AM EDT – ‘Collective Intelligence’
08:12AM EDT – Individual Intelligence’
08:11AM EDT – Human beings are always the source of intelligence
08:11AM EDT – But we need more intelligence
08:11AM EDT – Using a smartphone for more and more, for work and everything
08:10AM EDT – Innovation is growing so fast
08:10AM EDT – Today I want to show you Huawei’s approach to AI with our new AI processor
08:10AM EDT – Announcing Richard Yu
08:09AM EDT – Huawei employs 40k, as much as IFA gets in visitors per day
08:09AM EDT – ‘Huawei is one company investing in AI technology’
08:09AM EDT – ‘For a thermostat, you can wait a millisecond to compute what is needed in the cloud. In a car, you can’t have that latency and the compute needs to be local’
08:08AM EDT – ‘AI requires some serious lifting for computing power’
08:08AM EDT – ‘In just a few years we will live in an AI first world, not a smartphone first world’
08:07AM EDT – ”AI has been talked about for decades, but only now are they ready for consumers’
08:07AM EDT – IFA Exec Director on the stage up first, to announce Richard Yu. But first a little about IFA
08:06AM EDT – People sitting on the floor ahead of the front row
08:06AM EDT – OK here we go
08:05AM EDT – Still waiting. Some people are wandering to the front row expecting a seat
08:01AM EDT – Still waiting to start. Again, our Kirin 970 piece is worth reading just before we start 🙂 http://www.anandtech.com/show/11804/huawei-shows-unannounced-kirin-970-at-ifa-2017-dedicated-neural-processing-unit
07:59AM EDT – Perhaps I should learn more Chinese. The conversation would be interesting
07:58AM EDT – I’m sitting in the front row, next to people I believe to be executives from ZTE.
07:56AM EDT – We’re a few minutes from starting, and the hall is full. People are standing in the wings
07:55AM EDT – In our piece I said it was likely 4xCA with MU-MIMO and 128-QAM. I think our readers got a better read on the situation, suggesting 256-QAM instead (because 128-QAM doesn’t exist as a standard)
07:54AM EDT – The integrated LTE is going to be supporting Category 18, good for 1.2 Gbps download
07:54AM EDT – The last sticker item for the Kirin 970 (that we know about) is the modem
07:54AM EDT – It will be interesting to see how the G72 power consumption comes out. We saw big power spikes with Huawei’s G71 implementation. They should be more in tune with G72 this time around, we hope
07:53AM EDT – That makes moving from an MP8 to an MP12 configuration easier too, which would traditionally offer a direct 50% peak performance increase in GPU loads
07:53AM EDT – This is because it was difficult to fully use that second pipe, so by reducing it to just an ADD, it reduces die area without a decrease in performance
07:52AM EDT – The difference between G71 to G72 is in the pipes: G71 uses two FMAD pipes per GPU sub-core, whereas G72 converts one pipe to just an ADD
07:51AM EDT – The 970 is however moving from a MaliG71 MP8 to MaliG72 MP12
07:51AM EDT – The other side of the coin is that it’s still very early for A75/A55. ARM is expecting these cores to see the light of day next year, probably later than earlier (depending on how adventurous some designers are)
07:51AM EDT – Taking a known core and shrinking the node is easier than taking a new core and putting it on a new node
07:50AM EDT – On the core specifications, Huawei is keeping 4xA73 and 4xA53 in the design. This is for two reasons
07:49AM EDT – they showed image recognition at a rate of 2000 pictures a minute (33 per second), compared to 97 pictures per minute using just the CPU
07:49AM EDT – Huawei has some 5000 software developers apparently, so there is a goal for their own apps to leverage the NPU
07:48AM EDT – One of the NPUs troubles will be vendor lock-in. No app developers will take the time to learn how to use it unless it’s integrated into regular compute APIs, because if they learn to use it then it’ll lock the software to Huawei
07:48AM EDT – I’m told it’s a sizable chunk of the silicon die. I doubt we’ll find out today about that, but it’ll be interesting to see a die shot
07:47AM EDT – this is a dedicated ISP that can provide 1.92 TFLOPs of FP16 performance in neural networks
07:47AM EDT – Next on the list is the new dedicated neural processing unit (NPU)
07:47AM EDT – When Huawei launched the Kirin 960 last year, it was slightly too early for the first wave of 10nm, so they spent another generation of 16FF+ on their flagship. Now arguably 10nm is more mature, and ripe for the picking
07:46AM EDT – The key new features for the Kirin 970 start with the move to TSMC’s 10nm
07:46AM EDT – http://www.anandtech.com/show/11804/huawei-shows-unannounced-kirin-970-at-ifa-2017-dedicated-neural-processing-unit
07:46AM EDT – We posted our Kirin 970 news yesterday, with the limited details we had
07:40AM EDT – Just me for images and photos today. Today is the first public day of IFA, so there’s a lot of foot traffic in the halls. This keynote should be full, if not just with press
07:40AM EDT – Huawei has a keynote at IFA this year. Having quietly announced the Kirin 970 and its new Neural Processing Unit yesterday without a word through the regular press channels, we’re expecting to here Huawei’s future march into AI from Richard Yu, CEO of Huawei’s Consumer Business Group (CBG).
Autore: Ian Cutress AnandTech