edit| GTIC
core things reported on September 8 that during the just concluded GTIC 2022 Global AI Chip Innovation Summit, Houmo Intelligent Founder and CEO Wu Qiang delivered a speech on the topic of "Using the Integrated Storage and Computing to Help the Computing Power Revolution in Intelligent Driving" .
Wu Qiang believes that the integration of storage and computing is the golden key to unlocking computing power and power consumption problems. This innovative computing architecture has the characteristics of large computing power, low power consumption and low latency. It has weak dependence on the process and can use the 28nm process Create performance or energy efficiency ratios that can be achieved by 7nm or even 5nm processes based on traditional computing architectures.
In the first half of this year, Houmo Intelligent's first integrated memory chip was successfully lit up, meeting the computing power and accuracy requirements required for the complex AI scenarios of intelligent driving for the first time. It successfully ran 11 categories of typical intelligent driving algorithms, with energy efficiency ratios up to 14% 20TOPS/W.
Wu Qiang revealed that his first chip is expected to be delivered to some customers for trial at the end of this year or early next year.
The following is the transcript of Wu Qiang's speech:
Hello everyone, I am very happy to represent Houmo Intelligent to attend this summit. The theme of my speech today is "Using the integration of storage and computing to help the computing power revolution in intelligent driving." There are two keywords in
. The first is storage and computing integration, and the second is intelligent driving. Let me start with intelligent driving.
1. Intelligent driving is popular and commercially available, putting forward new requirements for chips
The automobile industry is undergoing a major change that has not been seen in a century, mainly new energy and intelligence, which will bring great results to people's travel and lifestyle Change, in a sense, this change is no less than when cars first appeared more than 100 years ago.
In today's China, we are at the forefront of intelligence and electrification. We can already feel this change. For example, in most cities today, if you get on an online car-hailing car, it is very likely that this is a new energy vehicle , You will hear the driver tell you that new energy vehicles are much cheaper than fuel vehicles, and you have a lot of money in your pocket and are very happy.
Whether it is a new energy vehicle or a traditional fuel vehicle, automotive intelligence is a major trend, and this trend of intelligence is faster than we imagined. For example, many experts expect that the intelligent penetration rate may reach 25% in 2023, but in fact, our intelligent penetration rate of L2 and above this year has reached nearly 30%, and may reach nearly 50% by 2025. This means that there are 25 million cars in China every year, and about 10 million cars will be smarter than L2 or above.
This kind of intelligence has also changed our lives. Many of them are a kind of urgent need. A few days ago, a friend told me that because he had something at home, he had to drive home 1,000 kilometers one way, and one weekend, 2,000 kilometers on the expressway. It is very easy to have smart assisted driving.
The popularization and commercialization of intelligent driving has put forward many new demands for chips. There are three elements to popularize intelligentization: First, it is strong intelligence, which means that this thing is smart enough, at least it makes me very easy to drive 1,000 kilometers on the highway at least. ;The second is low cost, because most cars in China are between RMB 100,000 and RMB 200,000, and these cars can be afforded; the third is safety, and life is at stake in the use of cars.
From the chip perspective, it corresponds to 's large computing power, low power consumption, low cost, and high reliability .
From the perspective of computing power, there is a core contradiction between the development of smart chips to this day.
2. The core contradiction in the development of intelligent driving chips
on the demand side, and the demand for computing power is growing.
For example, I compared Tesla Model 3 with NIO cars that just released this year: First, there are a lot more sensors and a lot more pixels, 1.2 million to 8 million are higher, which will correspond to The computing power is greatly enhanced; second, the algorithm has become more complex and is still evolving, which means that I need the chip to have more open support for algorithms, because it must be able to support future algorithms.
What is the current approach?
First of all, I want to make a lot of computing power. In the past, it was dozens of TOPS, but now it is hundreds or even thousands of TOPS.In smart driving chips, I use advanced processes, from 7nm to 5nm or even to 3nm, and I use HBM to increase the storage bandwidth.
But these things, first, are very expensive and only big manufacturers can afford them. For example, HBM is the best Nvidia and AMD. Another method is like using soft and hard to couple to increase efficiency, which improves efficiency but sacrifices versatility.
So in a sense, the core contradiction we face in today in intelligent driving computing is:
first, it requires a lot of computing power, but it is very costly to do it in the traditional way.
Second, large computing power must represent high power consumption, because it is almost 1TOPS/W, and air conditioners can be used in the data center. I can't dissipate heat in the car, so it will cause great cost and stability to dissipate heat in the car. Sexual problems.
Finally, a solution for decoupling by soft and hard is necessary. This is a core contradiction facing smart driving chips today.
3. Integrated storage and computing, the golden key to unlocking computing power and power consumption problems
. In all methods, storage and computing are integrated, which is an innovative computing architecture. Among all the new technologies, it may be a golden key that can solve our computing power and power consumption today.
From the traditional von Neumann to today's storage and computing integration, in the strict sense, it is not near-realization. I understand that it is an improvement. It only partially solves the problem, but only the real storage and computing integration is a real storage and computing integration. Only by solving this problem from the bottom level.
The benefits it can bring are: it can make computing power larger, and in addition, it has the advantage of energy efficiency ratio, which can avoid the handling of big data. It can extend the energy efficiency ratio by an order of magnitude, and its extension There are also improvements. There is another most important point. Many people do not realize that in today's international environment, we are facing technological blockade, and the dependence of storage and computing integration on technology is relatively weak. It can use the more traditional 28nm process to make others 7nm. Even the performance or energy efficiency ratio of 5nm is its inherent advantage.
But it also faces challenges, after all it is a new thing. I have always compared it, just like when electric cars first appeared, there were many design challenges, because fuel cars have been made for hundreds of years, and electric cars have just begun. Because von Neumann has been working on it for many years, everyone knows how to design it, and how to achieve greater computing power in storage and computing is just beginning.
postmotor chooses to use a memory-computing integrated chip to do edge scenarios such as intelligent driving, unmanned vehicles, bionic robot , etc. This itself is a relatively perfect match from technology, product and market.
So how can we turn the technical advantages of storage and computing into a product advantage, and solve the specific pain points of downstream customers.
Let me give you a specific case. Many mid-range cars care a lot about the cost, and heat dissipation is actually a big problem. There are three ways: one is natural air cooling, the other is fan, or liquid cooling. Natural heat dissipation is the most ideal in terms of cost and reliability, but there is a mandatory requirement, which is that the power consumption is less than 15 watts.
There is no problem with traditional NCUs, everyone is naturally dissipating heat. Intelligent driving has brought a new headache to car manufacturers. I have never encountered such hot chips. How to dissipate heat, whether it can be used to cool liquids and increase costs, is a very concerning issue for car manufacturers.
If it is in the traditional von Neumann architecture, it looks like 1 watt and 1 TOPS, and 15 watts and 20 TOPS or higher, which actually cannot meet the needs of L2++ or higher autonomous driving.
and the integrated computing power of storage can use the characteristics of high energy efficiency, such as 15 watts of power consumption up to 60TOPS (physical computing power), or higher computing power, meeting the needs of L2++ intelligent driving.
4. The first integrated storage and computing chip is lit up this year, and it has successfully completed the mainstream intelligent driving algorithm
. The rear motorcycle is mainly two groups of people, one group of people is a big bull in integrated storage and computing, and the other group of people It is a group of people like me who have been working in the industry for 20 years, are making big chips, CPU, GPU, or smart chips.
We two groups of people collided together. How to solve the problems we are facing? We thought of these paths that are more recognized and optimistic about the integration of storage and computing to make intelligent driving chips and large computing power chips, so we came together and established this company.
Houmo has also been recognized by first-tier capital. It currently has R&D centers in Nanjing, Shanghai, Beijing and Shenzhen. This year, we have also lit up the first chip with large computing power and have run through some mainstream algorithms for intelligent driving. .
uses memory and computing integration to make a large computing power chip. This is an AI chip. How to design circuits and how to integrate AI cores, such as how to integrate them at the architecture level, including compiler and algorithm levels. These are all faced by us since our establishment. The challenge is the core of our solution. Houmo was established for more than a year and has applied for about 30 patents.
Let me give you an example. This is the design of our first generation of AI cores. The whole design is our own design and adopts a layered distributed design. Starting from the right, it is the bottom layer. We call it Macro, which is a memory computing unit. It is a unit that does calculations in memory. According to different needs, it can have different sizes, such as 512×64, 64×64.
Several Macros can form a Macro Group, and multiple Macro Groups can form a Macro Group Array, and Array becomes a part of Tensor Engine. My Tensor Engine is mainly Macro Grou p.
A Tensor Engine forms a Tile, which is a bit like Core similar to FSD, and is an independent computing unit. Several Tiles can form an AI Core, and multiple AI Cores can form a Cluster. The Cluster becomes a SoC, which is the most important AI computing unit in automotive-specific SoCs.
Overall, a Tile itself is similar to an FSD Core, which is an independent computing unit. Multiple Tiles can also be combined to perform operations, and multiple AI Cores can also be combined to perform a joint operation. So overall it is a perfect combination of distribution and concentration, which is our first generation design idea.
Since we were established, we have successfully made the first chip last year. This chip is the first time in the industry to integrate storage and computing to achieve dozens of TOPS or even higher chips, with energy efficiency of about 20TOPS/W.
We lit it up and successfully ran through some mainstream intelligent driving algorithms, such as detection, recognition, and semantic segmentation.
This should be the first time in the industry to use the integrated memory chip to make an attempt to do mainstream AI application scenarios for intelligent driving. This is based on a mass production design. We have considered redundancy, self-repair, etc., and are completely based on A product-grade design. Our first-generation chips are expected to be delivered to some customers for trial at the end of this year or early next year. This is some progress in our products.
5. Future planning: Going towards a 100x energy efficiency ratio, 2000TOPS computing power
Let’s talk about our product planning in the next few years.
, the memory and computing are integrated. It is based on different storage media. The previous generations were basically based on mature SRAM. SRAM is now relatively mature and has good performance, but its capacity is limited.
We hope that when using SRAM, we can make a large computing power chip that can break at least 500 TOPS to 1,000 TOPS.
2.0 stage, our chip is based on more advanced storage media such as MRAM and RRAM. We hope that it has a larger capacity, can produce greater computing power and higher energy efficiency ratio, and can launch more competitive products.
Houmo's chip is based on the underlying architecture innovation, and it is unconscious for customers. He doesn't care how you achieve it. He doesn't care whether you use storage or von Neumann. He just needs to see a better one The products have better performance, better delay, etc., which are the results we hope.
is more important for Houmo. We try our best to connect with mainstream AI programming frameworks, including, for example, the application layer TensorFlow, PyTorch, ONNX, etc., and then we provide ourselves an intermediate layer to perform insensitive conversion.
We only do reasoning, hoping to have a very light tool to convert these inference models so that everyone can feel unconscious.
In addition, we are making relatively general computing chips. As I said before, the algorithm has been evolving, and everyone has requirements for openness, because computing power itself is an underlying architecture and can be very general. We can provide a series of applications. The computing architecture can also provide underlying interfaces, allowing customers to customize operators, which is a plan and what we are doing in software.
Overall, this is our software stack. Basically, we hope it is a design concept of software and hard decoupling. There are three colors here, which looks a bit more: brown is the hardware, chip, driver software, and SDK we provide; black We will provide a reference design, including reference hardware and solutions; gray is the customer's own thing, more of the application layer, including their operating system, middleware, etc., basically three layers.
We hope to provide customers with a reference design to decouple soft and hard to create their own things with the greatest flexibility. This is the design concept of Houmo Intelligent.
6. Vision: Create 1P/W computing power to support all things intelligence
Finally, share an interesting industry observation. I think in a country or region, the prosperity of an industry will surely drive or incubate leading enterprises in the local supply chain.
For example, in the 1970s and 1980s, people remember the prosperity of cameras and other excellent semiconductor companies such as Panasonic and Sony were born. By the 1990s, the prosperity of American PC products certainly also led to large semiconductor companies such as , Intel , and AMD.
The reason is very simple, because the product manufacturers and suppliers will have a virtuous cycle. The product manufacturers will give the demand to the suppliers, and the suppliers can make better components and better services, because they are local, When the manufacturer comes, he will give more orders to the supplier, and finally form a virtuous cycle.
So my own conclusion is that Chinese automobiles will definitely rise in electrification and intelligence, which is a trend. It will inevitably drive local supply chain companies, including chip companies, to become bigger and stronger.
So I also hope that our Houmo Intelligent and some of the brother AI chip companies here can seize this historical opportunity, become bigger and stronger, and become China's own international chip giant.
Finally, I will talk about Houmo Intelligence's vision. The picture on the left was extracted from the speech of Academician Liu Ming of . She told an interesting thing, saying that every 1,000 times the chip energy efficiency ratio will be extended to a new one. The calculation form of the computer, such as the small computer era, the PC era, and the mobile phone era is about 1T/W. Her prediction is that if we really reach the era of intelligence in everything, we may need 1P/W of computing power and energy efficiency ratio.
I like Academician Liu Ming’s speech very much. I brought it in and expressed the original intention of Houmo Intelligent’s entrepreneurship. We hope to use more underlying technologies to create 1P/W computing power.
Houmo Intelligent initially used the intelligent driving of unmanned vehicles, robots, and passenger cars as the entry point, to implement the products and create value for customers.In the future, we hope to expand into fields such as bionic robots and VR/AR. In the end, we hope to build a computing power platform in the era of artificial intelligence to realize the beautiful vision of intelligence of all things.
or above is my sharing, thank you everyone.
or above is a complete compilation of Wu Qiang's speech.
