Only NPU can make smartphone more intelligent?

 

The NPU is one of the hottest technologies in the handset industry this year. Apple A11 Bionic, Huawei Unicorn 970, and the Image Processing Unit (IPU) inside Google Pixel 2 are typical examples. After equipped with NPU, the common point is that the phone has a learning ability, if slightly more complicated, that is, with the number of product algorithm and matrix multiplication, etc., the rapid implementation of complex calculations of mathematical functions. Then what presented to the user, of course, is faster, smarter mobile experience.

 

Of course, the integration of a single NPU, cell phone circuit design is more complicated, which is why only high-end mobile phones will be integrated with NPU. Although machine learning is the only way to intelligent hardware, some media think that machine learning does not necessarily require the phone's built-in NPU to achieve. ARM released the Cortex-A75 and A55 CPUs and Mali-G72 GPU designs earlier this year, all with more advanced machine learning algorithms, but not through the NPU.

 

NPU (Neural Processing Unit) has not been well-known, but it is a hot technology in the field of chips. In contrast to CPU processors in the von Neumann architecture, it uses a disruptive new architecture called "Data Driven Parallel Computing." If regard the Von Neumann architecture handles data as a single lane, then Data-Driven Parallel Computing is a 128-lane multi-lane parallel that can process up to 128 data at a time, making it easy to handle large amounts of multimedia data in video and image formats.

 

The real performance of Huawei mate10NPU

 

Unicorn 970 is Huawei's first artificial intelligence mobile computing platform, is also the world's first independent AI artificial intelligence NPU neural network processing unit chip, using the innovative HiAI mobile computing architecture.

 

On Huawei AI this feature mainly focus on the camera, automatic recognition of food fresh flowers text animals different subjects, depending on the subject to be optimized, and the CPU than Kirin 970 AI computing power up to 50 times the energy efficiency and 25 times performance. In NPU's blessing, under the same shooting scenes Kirin 970 shooting speed is also faster than the average CPU. Compared to iPhone7 Plus 487, equipped with unicorn 970 intelligent terminals within 1 minute can identify over 2000 pictures.

 

The unicorn 970 features an innovatively designed HiAI mobile computing architecture with an energy-efficient heterogeneous computing architecture that dramatically increases the computational power of AI, of which AI performance density significantly outperforms CPUs and GPUs, enabling AIs to complete computing tasks faster with less power consumption, which is completely different from server-side AI design.

 

Machine learning does not necessarily depend on NPU

 

In terms of the hardware structure of current smartphones, the NPU can only do math on 8-bit and 16-bit data instead of 32-bit and 64-bit. You know, this operation saves the memory cache requirements, but increased bandwidth requirements, so the actual effect is not ideal. Therefore, ARM believes that integrating a more advanced single instruction, multiple data architecture is more resource-saving. The new INT8 can combine multiple instructions into one instruction to improve the delay phenomenon.

 

In addition, ARM solutions can help reduce costs, although there is no custom SoC based on A75 / A55 release, Qualcomm, MediaTek, Samsung and Hisilicon are expected to take advantage of the above instruction setting improvements to develop better SoC, Including mid-range products.

 

The key to ARM's ability to drive SoC machine learning is the Compute Library, which includes a full suite of capabilities for imaging and visual projects and a machine learning framework such as Google's TensorFlow which developers can use to recompile the version what they need.

 

In addition to ARM, Qualcomm also has its own Hexagon SDK, which also includes machine-learned generic matrix algorithms for more efficient DSP operation. In addition, the Qualcomm Symphony System Manager SDK provides an API that specializes in optimizing computer vision, image and data processing, and low-level algorithms as well as the common processing needs of smartphones.

 

Why does NPU exist?

 

Since NPU is not necessary, and architects manufacturer such as ARM also provide other methods, why Apple, Huawei and Google insisted so? The answer is not simply to improve selling points, hardware prices, but in some ways NPU still has some advantages.

 

For example, the unicorn 970 FP16 instruction set throughput is 1.92TFLOP, which is 3 times of the Mali-G72 version. In the role of dedicating hardware acceleration, optimization, standalone NPU performances more powerful. Of course, these built on the condition of manufacturers know very well and have completely control of their own hardware, which means that manufacturers need to spend more time and cost to research and development, mobile phone prices will not come down.

 

It is estimated that by 2018, high-end handsets will still adopt a solution with a stand-alone NPU, but not everyone will buy a high-end smartphone like the Xiaomi Redmi Note 4X. Low-end smart phones will be more inclined to choose ARM's low-cost solutions to achieve better machine learning capabilities.