Home > News content

AI performance up to 480 times! Arm in a hurry?

via:驱动之家     time:2020/2/13 1:01:09     readed:1169

Yesterday, ARM unveiled two new IPs, a new member of the already highly successful ARM Cortex-M family, Cortex-M55, and a new line of Ethos late last year, the first micro-NPU for the Cortex-M platform

The significant improvement of AI performance of the two new products is a powerful weapon for arm to compete in the aiot market. But it also reflects arm's anxiety in this market.


[Picture: Techviral]

Attack the aiot Market

Starting with two new products, the new Cortex-M55 is ARMAI's most powerful Cortex-M processor. How powerful? Compared with previous generations of Cortex-M processors, the ML performance of Cortex-M55 can be increased by up to 15 times, and the performance of DSP can be increased by up to 5 times, and it has a better energy consumption ratio. At the same time, Cortex-M55 is the first processor based on ARMv8.1-M architecture, built-in ARMHelium vector processing technology, significantly improving the performance of DSP and ML and more power saving.

In addition, developers can also use the new functions of Cortex-M processor, that is, the arm custom instructions (arm custom instruction set) released last October to extend the processor's capabilities and optimize specific workloads.

In other words, the new cortex-m55 not only has stronger AI performance, better energy consumption ratio, but also has more flexibility.


Another new product, ethos-u55, has more powerful AI performance, 32 times that of cortex-m55, and 32-256 configurable Macs. The purpose of ARM to launch this special micro neural processor is to accelerate the ML reasoning ability of space limited embedded and Internet of things devices by virtue of its high configurability.

This is not enough. If the cortex-m55 is combined with the ethos-u55, the ML performance of the combined product can be increased by up to 480 times.


Of course, don't be surprised by the 480 times improvement, because this performance improvement is based on a relatively low ml performance. Use the following figure to intuitively explain the applicable scenarios after AI performance improvement.


Cortex-M55's AI performance can be applied in the applications of image anomaly detection and object detection, while Cortex-M55 and Ethos-U55 can be applied in the applications of gesture detection, biometric recognition and speech recognition with higher requirements for AI performance, but the AI performance of the two combinations is lower than that of Cortex-A CPU, Mali GPU, Ethos-NPU can meet the needs of object classification and real-time recognition application.

In this way, we can clearly understand the meaning of cortex-m55 and ethos-u55.

Cortex-M is a very successful microcontroller in the industry. Cortex-M-based chips have shipped more than 50 billion pieces worldwide and are widely used in a wide variety of embedded devices. But Cortex-M can only meet the simple AI functions of vibration detection, sensing fusion, keyword detection, and with the arrival of the age of all things AI, we need the performance of Cortex-M, Ethos-U55 and the combination of the two to meet the needs of more AIoT applications.

That is to say, the existing Cortex-M, cortex-m55, ethos-u55, and the combination of all Cortex-M products and ethos-u55 has more comprehensive AI performance, which improves the attraction of arm in terminal AI. In addition, the new function of Cortex-M's arm custom instructions can better meet the large but fragmented aiot market.

Internal and external anxiety of arm in aiot Market

From Cortex-M to Ethos-N, ARM should have felt confident about the AI products it has, but the competition in the market doesn't seem to allow ARM to be satisfied for that. In fact, ARM unveiled Project Trillion's machine learning plan in February 2018 and is expected to release its first ML processor in the first quarter of 2019. But ARM didn't release its first ML processor until May 2019, and it didn't release the name Ethos for the ML series in October.

Dennis laudick, vice president of business and marketing of arm ml business group, explained the delay in product release in an exclusive interview with Lei Feng last year. There are many reasons for the delay, and product development is the main problem. In the process of ML processor development, there are many challenges, one of which is how to balance efficiency and flexibility.

When arm released its first ml processor, the NPU integrated in its SOC by apple, Samsung and Huawei had been updated for several generations, and the neural network processors of Qualcomm and MediaTek were becoming more and more mature. Therefore, many people in the industry think that arm has missed a good opportunity to enter the mobile NPU market. This will naturally make arm's AI products pay more attention to the aiot market with broad prospects.

A forecast released last year by Marketsand Markets, the world's second-largest market researcher, showed that the global AIoT market was $5.1 billion in 2019, which would grow to $16.2 billion by 2024, with a compound annual growth rate of 26 percent.

Although ARM has been very successful in the embedded device market with its Cortex-M series processors, in the era of AIoT, ARM has had a strong opponent


Risc-v's open source can enable developers to achieve more flexible design and more easily meet the diversified needs of the aiot market, which is also a key factor for arm to feel nervous. Last October, arm also launched the arm custom instruction set to cope with the competition from risc-v.

In addition, risc-v is growing, and products for the aiot market are also being launched.Green Waves Technologies pioneered the RISC-V-based IoT ultra-low-power processor, which can achieve 20 times more energy efficiency than similar products in the market below 100 MW. The performance of the new generation of products launched in December 2019 is increased by 10 times and the power consumption is reduced by 5 times.

Lin Zhiming, general manager of crystal core technology, an important IP provider of risc-v, also said in an interview with Lei Feng. Com last year that it will take another half year to promote mass production, listing and integration of software and hardware, and that risc-v products will be put into the market more in the second half of 2019 and the beginning of 2020.

At the beginning of this year, sifive and CEVA announced their cooperation to design and create edge AI processors specific to the ultra-low power consumption field for the final market in batches, with sifive's risc-v CPU, CEVA's DSP core, AI processor and software as the center.

Therefore, due to the delay of arm ml product release caused by many factors and the growing strength of risc-v camp, arm will inevitably have some anxiety in the era of AI and aiot, and 480 times AI performance improvement is a key to alleviate this anxiety.

Can the ecological advantage continue?

When asked if it was competitive to launch the first eth OS in May last year? Dennis said that arm's success has always benefited from the ecosystem. For NPU, the challenge is not to be able to make its own NPU, but whether NPU is really useful in the market. We have heard feedback from partners that they want to be standardized on hardware, they don't want to support 15 different kinds of hardware, so they expect to have a standardized software platform to support.

In terms of building the AI competitiveness of arm, even if the product launch is delayed, the strong ecosystem built by arm over the years will still contribute to the success of its NPU. In addition, arm is also committed to creating a consistent experience. Cortex-M software development tool fully supports the latest release of cortex-m55 and ethos-u55. In addition, arm is also compatible with other development tool manufacturers, and also has the support of partners in algorithm, software and OS.


Better software and hardware integration and consistent development process are indeed the attraction of an AI ecosystem, and also the key for arm to gain more market share of aiot.

But this may not be the case for many applications of AIoT, as GreenWaves Technologies CEO Loic Lietar previously told Lei Feng (Public: Lei Feng): from greenwaves ’ point of view, our risc-v based 32-bit microcontrollers are oriented primarily to marginal computing requirements from end markets such as the internet of things. a device will run several applications repeatedly for a long time without installing and uninstalling different applications like a phone or computer, so the degree of dependence on software compatibility is not a problem.

This means that arm's strong ecology will continue to play a role in the AI and aiot era, but for the aiot market, it is not clear how attractive this ecology is.

Of course, there is no doubt that arm, as a very successful semiconductor company, has strong products and rich experience. In the new market of aiot, arm also has strong competitiveness.However, it is difficult for the aiot market to generate giants in the short term, and it is difficult for a certain processor to monopolize the market in the short term. No one can accurately predict who will win.

China IT News APP

Download China IT News APP

Please rate this news

The average score will be displayed after you score.

Post comment

Do not see clearly? Click for a new code.

User comments