Xiaobian compares the current technology and market development of the mainstream "AI processor" and lists their processes, architectures and applications in the form of tables.
Artificial intelligence is hot, domestic and foreign chip developers compete to release their own AI processor smart chips. Xiaobian compares the current technology and market development of the mainstream "AI processor" and lists their processes, architectures and applications in the form of tables.
Cambrian-1: Core of the Core
Cambricon-1A of Cambrian Technology is a deep learning dedicated processor chip (NPU). Its high-performance hardware architecture and software support mainstream AI development platforms such as Caffe, Tensorflow and MXnet. It is said to be the first commercially successful deep learning processor IP product in the world, which can be widely used in key areas of intelligent processing such as computer vision, speech recognition and natural language processing. CB Insights, the US authoritative media focused on the development of the artificial intelligence industry, recently released the global AI 100 list. Cambrian was selected for its deep learning dedicated processor and is the only AI hardware startup in mainland China.
Cambrian currently has three product lines: First, the intelligent terminal processor IP license, intelligent IP command set can be authorized to be integrated into mobile phone, security, wearable devices and other terminal chips, customers including domestic top SoC manufacturers, has begun to enter the market . Secondly, in the field of intelligent cloud server chips, as a PCIE accelerator card inserted in the cloud server, the customer is mainly a well-known server manufacturer in China. Third, the home smart service robot chip: starting from smart toys and intelligent assistants, the service robot has the ability to see and hear independently. Customers are all types of downstream robot manufacturers, and the launch of the products will be later than the smart cloud server chips.
Huawei Unicorn 970: Known as the world's first AI processor
Huawei claims that the Kirin 970 is the world's first artificial intelligence system-on-a-chip. The Kirin 970 chose a heterogeneous computing architecture to dramatically increase the computing power of the AI, with a built-in independent neural network processing unit. This dedicated hardware processing unit is said to be derived from the Cambrian NPU IP license and is dedicated to machine learning and general AI applications.
Kirin970 uses TSMC's 10ns chipset process technology (20% reduction in power consumption and 40% reduction in volume). The main specifications are as follows: 8 core CPUs (up to 2.4GHz), new generation 12 core GPUs (Mali G72MP12), Kirin NPU (1.92T FP16 OPS), Image DSP (512bit SIMD), Dual Camera ISP (with face, motion detection), HiFi Audio (32bit/384k), UFS 2.1, security engine (inSE&TEE), global 4.5G modem (1.2Gbps@LTE Cat 18), 4K video (HDR10), LPDDR 4X, i7 inductive processor.
Huawei announced that the Huawei Mate 10 Pro, which will be launched in the US, will be equipped with the Kirin 970 chip, and Huawei Honor V10 will also be adopted. In contrast, Qualcomm Snapdragon 845 also focuses most of its focus on AI, and the penetration rate will undoubtedly be higher than Kirin 970. The Snapdragon 845 supports many Android flagship smartphones, including Samsung, Sony, LG and Xiaomi. .
Unlike Huawei's Kirin 970 chip, Qualcomm does core optimization in a common platform. It does not have a separate neural network engine unit, but a more flexible machine learning architecture, distributed on each unit such as CPU, GPU, DPS, etc. Thereby, the individual processing units can be flexibly invoked for different mobile terminals.
For the difference between the two directions, Qualcomm believes that integration is more effective. However, Huawei believes that in view of the energy requirements of mobile phones, the independent NPU processing unit must be the only way for mobile phone processors in the future. From now on, only Apple and Huawei have made independent NPUs.
Recently, the Kirin 970 and the Snapdragon 845 running from Weibo are compared, claiming that the former is 7% higher than the latter, but the difference between Kirin 970 and Xiaolong 845 is actually small, and the network only shows a few independent. The test results are not the average of the complete run points, and the display gap may even be smaller than it seems. Similarly, even if the processor has a high running score, the performance of the field operation is not necessarily the case, especially if the gap is so small. Having said that, the results of the leaked running points still suggest that the computing power of Huawei's flagship processor may soon catch up with Qualcomm.
Qualcomm Xiaolong 845: This year will be widely used in high-end Android phones
Snapdragon 845 uses the latest eight-core Kryo 385 custom architecture, performance is 25% higher than Snapdragon 835's Kryo 280, Samsung's second-generation 10nm process, up to 2.8GHz; followed by Snapdragon 845 integrated Adreno 630 GPU performance The Adreno 540 is 30% better than the Snapdragon 835 and consumes 30% less power. In addition, the Snapdragon 845 integrates the second-generation Gigabit LTE Modem, the X20 modem, which is 20% faster than the Snapdragon 835's X16. Its integrated Hexagon 685 DSP and Spectra 280 ISP fully enhance the camera function.
From the comparison of the specifications of the éªé¾™ 845 and Kirin 970 released by Weibo, the 845 core of the Snapdragon has undergone a huge upgrade. The new high-end large core A75 and A53 are combined, and the GPU is upgraded to Adreno 630 and Samsung 10nm LPE process technology. The Kirin 970 continues to use the A73 core and the A53 core. The GPU model is unknown. It also uses the 10nm process and is manufactured by TSMC.
Samsung Exynos 9810: Apple A11's strongest opponent?
Samsung's Exynos9810 processor uses its third-generation self-developed M3 architecture, with four 2.9GHz M3 cores and four 1.9GHz A55 cores, which are still 10nm (FinFET) process technology. 10nm is also the process technology currently used by Apple A11, Xiaolong 845 and Qilin 970.
On the GPU side, the Exynos 9810 uses the latest Mali-G72, which uses 18 cores (MP18) and is expected to operate at 700MHz. The Mali-G72 is a Bifrost-based graphics processor released by ARM last year, providing more powerful performance based on smaller area and lower power consumption. With the Mali-G72, the overall graphics performance is 1.4 times that of the previous generation. Energy efficiency increased by 25%, chip area performance increased by 20%, and machine learning efficiency increased by 17%.
In terms of artificial intelligence, Bixby, which supports face detection, has become smarter. Based on deep learning of neural networks, the new processor can accurately identify people or objects in photos by fast image search and classification for fast image search or classification, or scan the user's face in 3D by depth sensing. Perform mixed face detection. By using hardware and software, the hybrid face detection function can realize true face tracking detection, which makes it safer to use the face to unlock the device.
NVIDIA DRIVE Xavier: Powerful driving for autonomous driving
NVIDIA at the CES showcases DRIVE Xavier, a custom-built 8-core CPU, a new 512-core Volta GPU, a new deep learning accelerator, a new computer vision accelerator, and a new 8K HDR video. Built by the processor. DRIVE Xavier offers higher processing power, lower operating power, 30 teraflops per second, and power consumption of only 30 watts, 15 times more energy efficient than previous generation architectures. All AI computing tasks such as TensorCore, video recognition and stream processing, object location, path planning, etc. can be run fast, and the first batch of samples can be delivered to customers in the first quarter of 2018. NVIDIA CEO Huang Renxun said that the Chinese market is the largest market in the world. All systems are designed with localization and Chinese customers' needs in mind. For example, every auto-driving vehicle of Baidu is equipped with Drive Xavier.
Intel Loihi: challenging the neural mimetic computing puzzle
In order to defend against NVIDIA's strong offensive in the field of artificial intelligence, Intel has acquired FPGA chip giant Altera, AI startup Nervana Systems, and Israeli autopilot chip company Mobileye. At CES, Intel showed its self-learning neuromorphic chip "Loihi" to the public. This is a neuromorphic chip that was launched after the acquisition of the above technology companies and the collection of many research results.
AI chips can be divided into two categories, one is artificial neural network, and the other is neural mimetic calculation. In theory, the neural mimicry is more efficient, but the chip development is more difficult. Intel's Loihi uses neural mimicry to calculate this. The harder road can be seen as it wants to counter NVIDIA's ambitions.
Loihi uses architecture-to-chip modeling, asynchronous design flow, and round-robin algorithm verification based on FPGA simulation. It has very energy-saving features, 128 cores + 3 low-power X86 cores, and programmable learning rules.
Intel introduced that the Loihi chip has autonomous learning capabilities and can use data to learn and infer. It can become smarter over time and can be used in automotive and industrial production. But with so many advantages, it all needs to be In reality, to prove whether it has these capabilities, this is precisely the weakness of Intel compared to NVIDIA, because most of the current global neural network training systems use NVIDIA chips, Intel needs to increase the intensity to promote market share. On NVIDIA.
Rockchip RK3399Pro: The first CPU + GPU + NPU hardware structure design
Rockchip unveiled its first high-performance AI processor RK3399Pro at CES, providing a one-stop Turnkey solution for artificial intelligence with NPU performance up to 2.4 TOPs with high performance and low power consumption. , development and other advantages.
RK3399Pro AI chip adopts big.LITTLE size core CPU architecture, dual-core Cortex-A72+ quad-core Cortex-A53, with technical leadership in overall performance and power consumption; quad-core ARM high-end GPU Mali-T860, integrates more bandwidth compression technology, Excellent overall performance. RK3399Pro has strong AI computing performance. It is the first AI chip designed by Ruixin Micro with CPU+GPU+NPU hardware structure. Its integrated NPU combines Ruixinwei's years of experience in machine vision, speech processing and deep learning. . Compared with traditional chips, the typical deep neural network models such as Inception V3, ResNet34, and VGG16 perform well on the RK3399Pro chip.
MediaTek NeuroPilot: widely used in consumer products
MediaTek launched the NeuroPilot Artificial Intelligence (AI) platform at CES, focusing on terminal edge computing for smartphones, smart homes, and self-driving cars. MediaTek said that in the current year, about 1.5 billion consumer electronics products will use MediaTek chips. In 2018, AI processors and NeuroPilot SDK software development kit technology will be integrated to bring AI into a wide range of consumer products.
MediaTek has added the AI ​​core to the new Helio mobile phone chip in 2018. It has now built AI solutions for smart voice assistants, smart TVs and self-driving cars, and showcases cross-platform applications in CES, including Amazon Echo Smart Voice Assistant, Android. O Smart TV, BelkinWemo Smart Socket, and MediaTek cover the home router.
Huaxia core "North Star": AI chip platform with completely independent intellectual property rights
Huaxia Core released the AI ​​chip platform with full autonomous IP - "North Star", which is the first platform-based artificial intelligence chip with independent intellectual property rights for CPU, DSP and AI. "North Star" is a SoC chip for a variety of applications, not only AI dedicated processor for neural network and deep learning, but also integrated high-performance CPU / DSP, its ability can be extended to multiple product areas, such as intelligence Embedded artificial intelligence applications such as assisted driving, intelligent security monitoring, robotics, computer vision, on-board and commercial radar detection, and speech recognition. In addition, it can be extended to many other fields such as Industry 4.0, field control, edge computing, intelligent hardware, smart home, etc. It is a market-compatible heterogeneous computing and artificial intelligence platform chip.
"North Star" can use a programmatic extension method on a single chip to achieve high complexity of field control and decision making, digital signal processing, image signal processing, neural network-based deep learning and feature extraction, multi-threaded parallel computing and many other functions. . The "North Star" chip is manufactured by TSMC's 28nm process and will be mass-produced in the first half of 2018.
Horizon "Journey" and "Rising Sun": Embedded Artificial Intelligence Vision Chip
The Horizon, which received Intel's $100 million investment last year, launched Journey and Sunrise processors, all of which are embedded artificial intelligence vision chips for smart driving and smart cameras.
The performance of these two chips can reach 1Tops, and 1080P@30 frames can be processed in real time. Each frame can detect, track and identify 200 targets at the same time. Typical power consumption is 1.5W. The two chips use the data processing flow pattern of Attention Engine + Cognition Engine. Through this combination algorithm, the calculation speed of the chip can be more than 10 times. Through edge learning, the model can continuously improve itself, and the error rate is reduced by more than 50%. In addition, the two chips utilize an elastic tensor calculation core, and the multiplier utilization of the horizon artificial intelligence processor is close to 100%.
The journey 1.0 processor has the ability to accurately and timely detect and identify multiple types of targets, such as descriptions, motor vehicles, non-motor vehicles, lane lines, traffic signs, traffic lights, etc., and can support L2 level assisted driving systems.
The Rising Sun 1.0 series processor combines deep learning algorithms to support large-scale face detection tracking and video structuring in the front segment, which can be applied to smart cities and smart businesses.
Led Sign,Led Display Sign Board,Customized Led Display,Customized Led Display
ShenZhen Megagem Tech Co.,Ltd , https://www.megleddisplay.com