The battle for the landing of domestic smart driving chips in 2023 is imminent, and 'Post-Mo Intelligence' has revealed its killer app
In an era of burgeoning demand for computing power, storage and computing technology has gradually reached the eve of its explosion on the ground.
This technology was born in 1969, when the von Neumann architecture was leading the modern computer era and was the fundamental guideline on which the development of computers was "based". The concept of storage and compute as one, a disruptive fusion of the compute and storage units of a chip, was a fundamental solution to the potential energy efficiency bottleneck of separating compute and storage in the von Neumann architecture.
However, the technology relied heavily on storage media, and the process and performance gap between memory and processor was not obvious in the early days, so for decades after its birth, storage and compute remained a lab technology.
It was not until around 2018 that Moore's Law was approaching its physical limits due to the explosive growth of artificial intelligence applications and the big data industry, while new storage media and other storage and computing technologies were making breakthroughs. As the energy efficiency bottleneck of the traditional von Neumann architecture becomes more and more prominent, the storage and computing technology finally has a chance to get a ticket to commercialisation and become one of the highly promising technology trends in the post-Moore era.
Against this backdrop, a number of companies in China have been moving along the path of storage and computing integration technology in recent years. However, at this stage, most of the domestic storage and computing chips are mainly for end-side applications such as intelligent voice, which require relatively little computing power.
Houmo Intelligence is one of the early practitioners of the storage and computing industry in China. Since its inception in 2020, Houmo Intelligence has been aiming at the future era of intelligence in everything, providing the industry with the ultimate efficiency of storage and computing AI chips through underlying architecture innovation. After more than two years of "wild ride", Houmo Intelligence has finally delivered its first answer.
In the afternoon of May 10, Houmo Intelligence officially launched its first mass-produced storage and computation AI chip product - Hongtu™ H30, with 256T physical computing power and a typical power consumption of only about 35W. "Benefiting from the innovation of the underlying architecture, this should be the most efficient large computing power chip on the market. smart driving chip." Wu Qiang, founder & CEO of Houmo Intelligence, said with certainty.
That's right, the first trump card unveiled by Houmo Intelligence targeting the intelligence of everything is a smart driving chip, which is aimed at the autonomous driving market.
Perhaps in some people's view, the smart driving chip in 2023 is already in the critical stage of scrambling to get on board, and at this time to throw a new product to the market, after the Mo Intelligent whether to catch the landing window? Challenges abound in the industry environment, for Houma Intelligent, this is undoubtedly a key self-driving chip market battle for position.
01. The Hind Mo Hong Tou™ H30 was born as a result of the smart driving chip
Based on the storage and computing integration, the reconfiguration of intelligent driving chip is the first stage of Houmo Intelligent's goal for the future of everything intelligent, and is also the early strategic focus direction.
There is no other reason than demand is the most crucial decisive factor. In Wu Qiang's view, people on average spend nearly 1/8 of their day in the driving scene, smart driving will undoubtedly become one of the most important components of future intelligent life, "the end game of smart driving is to replace human driving, with sensors instead of eyes, with a variety of algorithms to fill the consciousness and soul, and the underlying smart driving chip is playing the role of the human brain. " he said.
From this perspective, a smart driving chip must require a computing method and efficiency infinitely close to that of the human brain, and the high computing efficiency of the storage and computing all-in-one has far exceeded that of traditional chip architectures, a technical advantage that is highly compatible with the key needs of future smart driving chips.
With this opportunity, the Hongtu™ H30 chip was born.
As the first storage and computation AI chip from Houmo Intelligence, the H30 chip uses a 12nm process and has a maximum physical computing power of 256Tops at Int8 precision with a typical power consumption of less than 35W. The H30 SoC based on the integrated storage and computing architecture has an energy efficiency ratio of 7.3 Tops/W, featuring high computational efficiency, low computational latency and low process dependency.
In fact, the H30's differentiating features are actually the advantages brought about by the underlying architectural innovation.
To put it simply, in the past, the efficiency of traditional architecture chip products mostly used the deep coupling of algorithms and chips, curing the algorithms into the chips, which to a certain extent lost the versatility of the chips. Based on the integrated storage and computing technology, the H30 is able to innovate from the underlying architecture to achieve a more cost-effective efficiency improvement without losing versatility.
In other words, although H30 is a relatively generic smart driving chip, it not only doubles the performance and computational efficiency and supports more algorithm models than most other dedicated smart driving chips, but at the same time consumes only one-half of their power.
To ensure the core competitiveness of the H30, improve the ease of use of the product, lower the migration threshold for customers and accelerate implementation, Houma Intelligent has also launched two products based on the H30: the smart driving hardware platform, Force Harness, and the software development tool chain, Houma Avenue.
Among them, the force harness platform as a domain controller, mainly for the end logistics unmanned small car, passenger car intelligent driving, vehicle-road cooperation and other scenarios. Its CPU arithmetic power is 200Kdmips, AI arithmetic power is 256Tops, supports multi-sensor input and further improves system reliability. At the same time, the Power Harness platform consumes only 85W of power and supports flexible heat dissipation methods, which greatly reduces the cost of convenient deployment.
The Hind Mo Avenue supports Pytorch, TensorFlow, ONNX and other mainstream open source frameworks, programming is compatible with CUDA front-end syntax, and supports both SIMD and SIMT programming models. The non-intrusive and innovative design of the underlying architecture also makes the H30 more efficient and easy to use.
From a comprehensive point of view, the two products, Power Harness Platform and Houmo Avenue, are more of a solution to the ease of use and rapid migration of H30. With the assistance of these two products, H30 will also achieve faster application implementation, thus further promoting the popular application of large computing power intelligent driving scenarios.
02. A better solution for storage and computing in the new battlefield of autonomous driving
The launch of a smart driving chip is just the beginning of a long and demanding validation cycle. In fact, 2023 is also known as the key year for domestic smart driving chips to hit the ground, and Houmo Smart chose to launch H30 at this point, which is a thought-provoking entry point.
"Is it late?" -- Inevitably, some people have doubts about the timing of H30's entry into the market, but Wu Qiang doesn't think so. He told 36 Krypton, if the market vein of smart driving chips is further subdivided, in fact, the real domestic replacement at the moment is only the L1 and L2 autonomous driving scenarios with low arithmetic demand, this track has long completed the process of domestic replacement from nothing to something, and the next step to take is the road from something to something better.
But in L2++ new battlefield, especially tens of T or even hundreds of T large arithmetic needs of the field of advanced autonomous driving, the domestic alternative has not really completed the first stage from nothing to have, and has not achieved a large number of landing of large arithmetic chip products, the localization options are still very limited. "The pattern of new battlefield players has not yet been determined, and it is not too late for H30 to enter the game." Wu Qiang said.
In Wu Qiang's view, in the face of today's new battlefield, local players to improve the rate of localization of smart driving chip breakthrough key lies in two points, one is the actual application performance of the chip to quickly reach the L2++ level of international giants; second is the software supply chain must be convenient and easy to use, so that customers can achieve rapid migration.
Along this track, which is swarming with players, Wu Qiang is confident, because Houmo Intelligence has chosen a different technical route, which is its biggest technical card - storage and computing integration. In fact, for traditional architectures, storage and computing integration is a new and complementary technology route that can achieve unique technological advantages in the post-MoR era where manufacturing processes and supply chain security requirements are becoming increasingly stringent.
As AI computing continues to evolve, the two core issues that will face future large-calculus AI applications are the bottleneck of data bandwidth and power consumption. As mentioned earlier, storage and computing in one can solve the increasingly prominent storage wall problem of traditional von Neumann architecture from the underlying architecture and break the energy efficiency bottleneck.
To use an analogy, the chip is like a huge factory, which needs to move data away from memory when performing each computing task, transport it to the processor through the bus data transfer channel, and then carry it back to memory when the processing is complete. This means that the processing speed of the chip is closely related to the bus speed, and once the amount of data to be processed gets larger, the computational efficiency will be constrained.
The core of storage and computation is the tight integration of storage and computation, in order to reduce the energy consumption and latency caused by the frequent transport of data between the two, and significantly improve the energy efficiency of the chip to achieve higher computational efficiency.
But this road is not easy. Wu Qiang bluntly said that before the birth of H30, Houmo Smart did a lot of internal testing and validation, "we have trickled through many potholes over the past two years, constantly upgrading and iterating the technology before handing over the first answer to customers."
In terms of results, six major technological breakthroughs were achieved behind the birth of the H30: large computing power, full accuracy, low power consumption, automotive-grade, mass production-ready and versatility.Chen Liang revealed that the second generation of Tianxuan architecture is now under development and will adopt a mesh interconnection structure, which will allow the number of computing units to be configured according to different application scenarios. The overall performance, efficiency and flexibility will further leap forward and will be geared towards cost and power sensitive smart terminals, large models and other scenarios. Planning has already started for the third generation of the Tiangui architecture, which is geared towards the intelligence of all things.
03. After the smart driving chip has landed, the storage and computing integration is accelerating fast towards tomorrow
As Zhang Yongwei, Vice Chairman and Secretary General of the China Electric Vehicle Council of 100, said at the launch of Houmo Smart's first storage and calculation integrated smart driving chip, China's automotive chip market is in the scale up period in 2023, and the chip demand is expected to be around 100 billion pieces, while the current global supply chain situation also provides a rare development opportunity and time window for domestic automotive chips.
"China's mega-market offers room for electrification in every technology route, and so does the development of chips." Zhang Yongwei said that the storage and calculation of integrated technology not only gets rid of the supply risk of advanced processes, but also pushes domestic players to further enhance the resilience and security of the supply chain through a variety of technology paths for localization layout.
And in the fierce globalisation race, landing is the only criterion to test product performance.
At this stage, the commercial landing of H30 is mainly for two scenarios: commercial vehicles and passenger cars. Xin Xiaoxu, co-founder and vice president of product of Houmo Intelligence, revealed that the company has successfully run the commonly used classic CV network and various advanced networks for autonomous driving based on H30, including the BEV network model, which is currently the most popular in the industry, and the Pointpillar network model, which is widely used in the field of advanced assisted driving.
The smart driving solution built with the H30 is now being deployed in partners' unmanned minivans, "the first time the industry has successfully run an end-to-end smart driving technology stack based on an integrated storage and computing architecture chip." Xin Xiaoxu said.
Not only that, H30 will start sending tests to Alpha customers in June. Meanwhile, the second generation of Houmo Smart, the Hongtu H50, is already in full development and is expected to be launched in 2024 to support customers' 2025 mass production models.
"L2++ is a necessary stage for fully automated driving, but it will definitely not be the end." Talking about the future, Wu Qiang told 36 Krypton firmly. He believes that fully automated driving will definitely collect more data through the continuous popularity of L2++, which will feed the development and popularity of L3 and L4 technologies, gradually and truly freeing people from the heavy scenarios to go to greater value.
Based on this, the path and direction of future product iterations of Houmo Smart is also becoming clear. The first is the arithmetic dimension, which will develop towards higher arithmetic power and energy efficiency; the second is the richness and integration of products, which will gradually cover more scenarios beyond the value of arithmetic power.
"Focusing on medium and large arithmetic scenarios, we will also launch a richer product range for different models and cost considerations to meet the demand of different types of customers for cost performance." Wu Qiang said.
Returning to the technology itself, if the timeline of the development of Everything Smart is further stretched, in fact the Smart Drive chip is not the only vehicle for the post-mo Smart technology to be implemented. This means that the future of Houmo Smart will also be in the track outside the smart driving chip to achieve multi-blossom, and the foundation of all this, can not be separated from the years of Houmo Smart on the storage and computing technology polishing, accumulation and innovation.
After more than 50 years of development, the all-in-one storage and computing technology has come to the eve of its formal explosion. The unveiling of Hongtu™ H30 officially opens a new phase of the commercial implementation of the domestic storage and computing all-in-one AI chip. Its birth is like a butterfly on the wing, and every step of its technological innovation and every loop of ecological layout will drive the future together with the industry players towards the day when intelligence in everything really comes.
This technology was born in 1969, when the von Neumann architecture was leading the modern computer era and was the fundamental guideline on which the development of computers was "based". The concept of storage and compute as one, a disruptive fusion of the compute and storage units of a chip, was a fundamental solution to the potential energy efficiency bottleneck of separating compute and storage in the von Neumann architecture.
However, the technology relied heavily on storage media, and the process and performance gap between memory and processor was not obvious in the early days, so for decades after its birth, storage and compute remained a lab technology.
It was not until around 2018 that Moore's Law was approaching its physical limits due to the explosive growth of artificial intelligence applications and the big data industry, while new storage media and other storage and computing technologies were making breakthroughs. As the energy efficiency bottleneck of the traditional von Neumann architecture becomes more and more prominent, the storage and computing technology finally has a chance to get a ticket to commercialisation and become one of the highly promising technology trends in the post-Moore era.
Against this backdrop, a number of companies in China have been moving along the path of storage and computing integration technology in recent years. However, at this stage, most of the domestic storage and computing chips are mainly for end-side applications such as intelligent voice, which require relatively little computing power.
Houmo Intelligence is one of the early practitioners of the storage and computing industry in China. Since its inception in 2020, Houmo Intelligence has been aiming at the future era of intelligence in everything, providing the industry with the ultimate efficiency of storage and computing AI chips through underlying architecture innovation. After more than two years of "wild ride", Houmo Intelligence has finally delivered its first answer.
In the afternoon of May 10, Houmo Intelligence officially launched its first mass-produced storage and computation AI chip product - Hongtu™ H30, with 256T physical computing power and a typical power consumption of only about 35W. "Benefiting from the innovation of the underlying architecture, this should be the most efficient large computing power chip on the market. smart driving chip." Wu Qiang, founder & CEO of Houmo Intelligence, said with certainty.
That's right, the first trump card unveiled by Houmo Intelligence targeting the intelligence of everything is a smart driving chip, which is aimed at the autonomous driving market.
Perhaps in some people's view, the smart driving chip in 2023 is already in the critical stage of scrambling to get on board, and at this time to throw a new product to the market, after the Mo Intelligent whether to catch the landing window? Challenges abound in the industry environment, for Houma Intelligent, this is undoubtedly a key self-driving chip market battle for position.
01. The Hind Mo Hong Tou™ H30 was born as a result of the smart driving chip
Based on the storage and computing integration, the reconfiguration of intelligent driving chip is the first stage of Houmo Intelligent's goal for the future of everything intelligent, and is also the early strategic focus direction.
There is no other reason than demand is the most crucial decisive factor. In Wu Qiang's view, people on average spend nearly 1/8 of their day in the driving scene, smart driving will undoubtedly become one of the most important components of future intelligent life, "the end game of smart driving is to replace human driving, with sensors instead of eyes, with a variety of algorithms to fill the consciousness and soul, and the underlying smart driving chip is playing the role of the human brain. " he said.
From this perspective, a smart driving chip must require a computing method and efficiency infinitely close to that of the human brain, and the high computing efficiency of the storage and computing all-in-one has far exceeded that of traditional chip architectures, a technical advantage that is highly compatible with the key needs of future smart driving chips.
With this opportunity, the Hongtu™ H30 chip was born.
As the first storage and computation AI chip from Houmo Intelligence, the H30 chip uses a 12nm process and has a maximum physical computing power of 256Tops at Int8 precision with a typical power consumption of less than 35W. The H30 SoC based on the integrated storage and computing architecture has an energy efficiency ratio of 7.3 Tops/W, featuring high computational efficiency, low computational latency and low process dependency.
In fact, the H30's differentiating features are actually the advantages brought about by the underlying architectural innovation.
To put it simply, in the past, the efficiency of traditional architecture chip products mostly used the deep coupling of algorithms and chips, curing the algorithms into the chips, which to a certain extent lost the versatility of the chips. Based on the integrated storage and computing technology, the H30 is able to innovate from the underlying architecture to achieve a more cost-effective efficiency improvement without losing versatility.
In other words, although H30 is a relatively generic smart driving chip, it not only doubles the performance and computational efficiency and supports more algorithm models than most other dedicated smart driving chips, but at the same time consumes only one-half of their power.
To ensure the core competitiveness of the H30, improve the ease of use of the product, lower the migration threshold for customers and accelerate implementation, Houma Intelligent has also launched two products based on the H30: the smart driving hardware platform, Force Harness, and the software development tool chain, Houma Avenue.
Among them, the force harness platform as a domain controller, mainly for the end logistics unmanned small car, passenger car intelligent driving, vehicle-road cooperation and other scenarios. Its CPU arithmetic power is 200Kdmips, AI arithmetic power is 256Tops, supports multi-sensor input and further improves system reliability. At the same time, the Power Harness platform consumes only 85W of power and supports flexible heat dissipation methods, which greatly reduces the cost of convenient deployment.
The Hind Mo Avenue supports Pytorch, TensorFlow, ONNX and other mainstream open source frameworks, programming is compatible with CUDA front-end syntax, and supports both SIMD and SIMT programming models. The non-intrusive and innovative design of the underlying architecture also makes the H30 more efficient and easy to use.
From a comprehensive point of view, the two products, Power Harness Platform and Houmo Avenue, are more of a solution to the ease of use and rapid migration of H30. With the assistance of these two products, H30 will also achieve faster application implementation, thus further promoting the popular application of large computing power intelligent driving scenarios.
02. A better solution for storage and computing in the new battlefield of autonomous driving
The launch of a smart driving chip is just the beginning of a long and demanding validation cycle. In fact, 2023 is also known as the key year for domestic smart driving chips to hit the ground, and Houmo Smart chose to launch H30 at this point, which is a thought-provoking entry point.
"Is it late?" -- Inevitably, some people have doubts about the timing of H30's entry into the market, but Wu Qiang doesn't think so. He told 36 Krypton, if the market vein of smart driving chips is further subdivided, in fact, the real domestic replacement at the moment is only the L1 and L2 autonomous driving scenarios with low arithmetic demand, this track has long completed the process of domestic replacement from nothing to something, and the next step to take is the road from something to something better.
But in L2++ new battlefield, especially tens of T or even hundreds of T large arithmetic needs of the field of advanced autonomous driving, the domestic alternative has not really completed the first stage from nothing to have, and has not achieved a large number of landing of large arithmetic chip products, the localization options are still very limited. "The pattern of new battlefield players has not yet been determined, and it is not too late for H30 to enter the game." Wu Qiang said.
In Wu Qiang's view, in the face of today's new battlefield, local players to improve the rate of localization of smart driving chip breakthrough key lies in two points, one is the actual application performance of the chip to quickly reach the L2++ level of international giants; second is the software supply chain must be convenient and easy to use, so that customers can achieve rapid migration.
Along this track, which is swarming with players, Wu Qiang is confident, because Houmo Intelligence has chosen a different technical route, which is its biggest technical card - storage and computing integration. In fact, for traditional architectures, storage and computing integration is a new and complementary technology route that can achieve unique technological advantages in the post-MoR era where manufacturing processes and supply chain security requirements are becoming increasingly stringent.
As AI computing continues to evolve, the two core issues that will face future large-calculus AI applications are the bottleneck of data bandwidth and power consumption. As mentioned earlier, storage and computing in one can solve the increasingly prominent storage wall problem of traditional von Neumann architecture from the underlying architecture and break the energy efficiency bottleneck.
To use an analogy, the chip is like a huge factory, which needs to move data away from memory when performing each computing task, transport it to the processor through the bus data transfer channel, and then carry it back to memory when the processing is complete. This means that the processing speed of the chip is closely related to the bus speed, and once the amount of data to be processed gets larger, the computational efficiency will be constrained.
The core of storage and computation is the tight integration of storage and computation, in order to reduce the energy consumption and latency caused by the frequent transport of data between the two, and significantly improve the energy efficiency of the chip to achieve higher computational efficiency.
But this road is not easy. Wu Qiang bluntly said that before the birth of H30, Houmo Smart did a lot of internal testing and validation, "we have trickled through many potholes over the past two years, constantly upgrading and iterating the technology before handing over the first answer to customers."
In terms of results, six major technological breakthroughs were achieved behind the birth of the H30: large computing power, full accuracy, low power consumption, automotive-grade, mass production-ready and versatility.Chen Liang revealed that the second generation of Tianxuan architecture is now under development and will adopt a mesh interconnection structure, which will allow the number of computing units to be configured according to different application scenarios. The overall performance, efficiency and flexibility will further leap forward and will be geared towards cost and power sensitive smart terminals, large models and other scenarios. Planning has already started for the third generation of the Tiangui architecture, which is geared towards the intelligence of all things.
03. After the smart driving chip has landed, the storage and computing integration is accelerating fast towards tomorrow
As Zhang Yongwei, Vice Chairman and Secretary General of the China Electric Vehicle Council of 100, said at the launch of Houmo Smart's first storage and calculation integrated smart driving chip, China's automotive chip market is in the scale up period in 2023, and the chip demand is expected to be around 100 billion pieces, while the current global supply chain situation also provides a rare development opportunity and time window for domestic automotive chips.
"China's mega-market offers room for electrification in every technology route, and so does the development of chips." Zhang Yongwei said that the storage and calculation of integrated technology not only gets rid of the supply risk of advanced processes, but also pushes domestic players to further enhance the resilience and security of the supply chain through a variety of technology paths for localization layout.
And in the fierce globalisation race, landing is the only criterion to test product performance.
At this stage, the commercial landing of H30 is mainly for two scenarios: commercial vehicles and passenger cars. Xin Xiaoxu, co-founder and vice president of product of Houmo Intelligence, revealed that the company has successfully run the commonly used classic CV network and various advanced networks for autonomous driving based on H30, including the BEV network model, which is currently the most popular in the industry, and the Pointpillar network model, which is widely used in the field of advanced assisted driving.
The smart driving solution built with the H30 is now being deployed in partners' unmanned minivans, "the first time the industry has successfully run an end-to-end smart driving technology stack based on an integrated storage and computing architecture chip." Xin Xiaoxu said.
Not only that, H30 will start sending tests to Alpha customers in June. Meanwhile, the second generation of Houmo Smart, the Hongtu H50, is already in full development and is expected to be launched in 2024 to support customers' 2025 mass production models.
"L2++ is a necessary stage for fully automated driving, but it will definitely not be the end." Talking about the future, Wu Qiang told 36 Krypton firmly. He believes that fully automated driving will definitely collect more data through the continuous popularity of L2++, which will feed the development and popularity of L3 and L4 technologies, gradually and truly freeing people from the heavy scenarios to go to greater value.
Based on this, the path and direction of future product iterations of Houmo Smart is also becoming clear. The first is the arithmetic dimension, which will develop towards higher arithmetic power and energy efficiency; the second is the richness and integration of products, which will gradually cover more scenarios beyond the value of arithmetic power.
"Focusing on medium and large arithmetic scenarios, we will also launch a richer product range for different models and cost considerations to meet the demand of different types of customers for cost performance." Wu Qiang said.
Returning to the technology itself, if the timeline of the development of Everything Smart is further stretched, in fact the Smart Drive chip is not the only vehicle for the post-mo Smart technology to be implemented. This means that the future of Houmo Smart will also be in the track outside the smart driving chip to achieve multi-blossom, and the foundation of all this, can not be separated from the years of Houmo Smart on the storage and computing technology polishing, accumulation and innovation.
After more than 50 years of development, the all-in-one storage and computing technology has come to the eve of its formal explosion. The unveiling of Hongtu™ H30 officially opens a new phase of the commercial implementation of the domestic storage and computing all-in-one AI chip. Its birth is like a butterfly on the wing, and every step of its technological innovation and every loop of ecological layout will drive the future together with the industry players towards the day when intelligence in everything really comes.
