Mysterious heterogeneous intelligent speech, energy consumption and calculation comparable to Google Huawei

Recently, the mysterious and heterogeneous intelligence of technology has finally appeared in people's sights. Wu Ren, CEO of Heterogeneous Intelligence, also expressed his vision of “the heterogeneous intelligence has chips, models, supercomputers, and hopes to get the model out of the cloud. Bring to the local and terminal, let the terminal become smart." Wu Ting talks about heterogeneous smart chips: energy consumption is lower than Google, higher than Huawei's Cambrian computing power.

The atmosphere and stability are the first impression left by Wu Ren. Since Baidu’s departure from the creation of heterogeneous intelligence in 2015, the CEO has rarely spoken out, and heterogeneous intelligence has been labeled as “mysterious”.

Rather than saying that the sharing of "Heterogenic Intelligence embraces the trend of artificial intelligence" is a media exchange meeting, it is better to say that Wu Ren reveals his talks about the current situation of heterogeneous intelligence.

"Heterogeneous intelligence has chips, models, and supercomputers. I hope to take the model out of the cloud and bring it to the local and terminal to make the terminal smart." Wu Ren expressed his vision for the company.

Mysterious heterogeneous intelligent speech, energy consumption and calculation comparable to Google Huawei

â–³ Novumind (heterogeneous intelligence) founder and CEO Dr. Wu Ren (right), and founder Liao Zhiyu (left)

Intelligent Terminal

"(Cloud) Internet of Things is a false proposition, it is difficult to achieve." Wu Ren said.

The current status quo is that the Internet of Things, which seems to be magically connected, actually needs to be connected to the main road from branch roads through many submarine and bottom cables. The branch roads that are accessed have grown exponentially and have far exceeded the growth rate of the main roads.

Because the traditional cloud architecture of the Internet of Things is difficult to process the massive amount of data in the future, heterogeneous intelligence begins to consider the ability of artificial intelligence to get from the cloud to the terminal, allowing edge devices to learn to "think." Connecting the results of thinking about edge devices to the world through the network is the realization of the intelligent networking (I2oT) in Wu Renkou.

"Doing artificial intelligence locally is the future direction, and heterogeneous intelligence will empower companies in this direction." Wu Ren said.

Putting the AI ​​on the terminal presents a two-point challenge to the chip. "High" is that AI chips are mainly powerful enough to support complex artificial intelligence calculations. "Low" means that the chip needs to be low enough to be placed on small objects at the edge.

"It's a challenge, and an opportunity," Wu Ren said with a firm voice. "The decision of heterogeneous intelligence is to take the challenge and seize the opportunity."

Mysterious heterogeneous intelligent speech, energy consumption and calculation comparable to Google Huawei

High performance - low energy consumption

In this scenario, heterogeneous intelligence chooses to abandon the versatility of the chip and design a dedicated chip for the AI ​​scene to pursue the highest performance-to-power ratio.

Wu Ren said that the current artificial intelligence chip really needs 10TOPS (teraflops per second) computing power, Google's top chip is high in computing power but high in energy consumption, and does not meet the conditions of local landing, and Qualcomm Although the chips such as the Cambrian Kirin 970 have low energy consumption, the computing power is not up to standard.

In contrast, the NovuTensor V1 can achieve 15TOPS performance with only 5W of power consumption. Low energy consumption and meeting computing needs.

Mysterious heterogeneous intelligent speech, energy consumption and calculation comparable to Google Huawei

â–³ NovuTensor advantage comparison

What is the concept of 5W? Wu Ren vividly stated that 5W is the power that USB can provide.

"This is the ability to break through the cloud to the terminal." Wu Ren said.

Patented full stack solution

"The core calculation of artificial intelligence is the convolution calculation of three-dimensional tensor, and the part of the small cube and the large cube overlaps," Wu Ren began to talk about the technology behind the chip. "The heterogeneous intelligent chip operates on the three-dimensional level. And there are four patent protections."

Mysterious heterogeneous intelligent speech, energy consumption and calculation comparable to Google Huawei

â–³ NovuTensor core technology

These four patents include direct implementation of hardware for three-dimensional tensor convolution, scheduling and distribution of data, merging, and collaboration between processors.

It is not enough to see the opportunity. Heterogeneous intelligence first found ways to design and implement chips and algorithms, supplemented by patent protection. In this way, heterogeneous intelligence has gone early and has stabilized.

"I want to empower other companies." Wu Ren mentioned this sentence more than once.

Empowerment, which is to help other companies train smarter models, more complex network structures, and make data sets play a bigger role.

By training for supercomputers customized for artificial intelligence, heterogeneous smart chips can implement design models with the highest performance-to-power ratios, providing full-stack services for different industries. It can even provide non-AI companies with the power of computing and intelligence to help artificial intelligence reach.

Mysterious heterogeneous intelligent speech, energy consumption and calculation comparable to Google Huawei

â–³ Heterogeneous intelligent full stack solution

future

The main smart modules of heterogeneous intelligence, NovuBrain, mainly include security, manufacturing, home, medical equipment, smart city and smart driving. Currently, we have cooperated with HP and KUKA.

"By giving technical capabilities to top-tier big companies that give these capabilities to their users, heterogeneous intelligence doesn't have to do these things in every industry, every user, every application." Say.

According to qubits, the current 28nm ASIC chip has entered the streaming stage, and the 16nm and 7nm chips are simultaneously advanced and are expected to be produced next year.

When asked if there is time to launch the 16nm chip next year, Wu Ren said that the company's team has accumulated a lot of talent and is confident in controlling the latest technology. The technical team's Mike Li was responsible for two 16nm chips in Inphi, and he has 100% confidence in its process.

"The future world must be the leader of the intelligent Internet of Things. The positioning of heterogeneous intelligence is to transfer the Internet of Everything to all things and provide all the technology for the transition. As the network grows up, the company will grow up and become a great company. Wu Ren said.

Portable battery

Portable Battery ,Portable Power Bank,Portable Battery Pack,Portable Power Pack

Zhejiang Casnovo Materials Co., Ltd. , https://www.casnovonewenergy.com

Posted on