Huang Xin, Zong Ren, Joint Editor.
NVIDIA today announced new products at GTC China 2016: Tesla P4 and Tesla P40, reasoning accelerators designed for deep learning, and Drive PX2 for autocruise, a low-power computer developed specifically for autonomous driving and automotive artificial intelligence.
NVIDIA's traditional strengths are desktop and mobile PC GPUs, but it's clear that it's moving toward AI is clearly not satisfied with simply improving GPU performance in a single area. We mentioned that AI's R&D and application are divided into several stages. NVIDIA's traditional computationally intensive GPU products will only significantly improve the training algorithm, but NVIDIA is obviously not willing to win. In this area, it is hoped that the entire development process will be included. P4 and P40 are similar to the P100 previously placed in the supercomputer DGX-1, but the functional focus is different. P100 focuses on accelerating the training speed of neural networks. P4 and P40 focus on improving the reasoning of the CPU, which is the logic operation efficiency. Here are the highlights of the 3 new products:
P4 VS P40 VS Drive PX2 Autocruise
1. P4 focuses on improving CPU reasoning, that is, logic operation efficiency.
Single precision floating point is 5.5 TeraFLOPS
The INT8 indicator is 22 TOPS
Built-in 8GB GDDR5 memory
2560 CUDA Core
Supports 192 GBPs of bandwidth.
Integrates 7.2 billion transistors
NVIDIA says P4 performance is four times that of the M4 released last year
Mainly responsible for image, text and speech recognition
2. P40 focuses on improving CPU reasoning, which is the efficiency of logic operations. CPU inference can be accelerated 40 times.
Single-precision floating-point 12 TeraFLOPS
INT8 indicator (measured for deep learning) is 47 TOPS
Built-in 24GB GDDR5 memory
3840 CUDA Core
346 GBps bandwidth
Integrates 12 billion transistors
NVIDIA says P40 performance is four times that of the M40 released last year
Mainly responsible for image, text and speech recognition
III. PX 2 AUTOCRUISE Computing Platform is suitable for auto-cruise function (including highway autopilot and high-definition drawing)
Real-time visibility of the situation
Precise positioning on high-precision maps and planning of safe driving routes. The platform uses a new single processor configuration
Helps vehicles use deep neural networks to process data from multiple cameras and sensors.
Power is only 10W
Focus on the Drive PX2 Autocruise
However, these are still NVIDIA's traditional strengths. Its data indicators are not only much faster and stronger than their predecessor, m4/m40, but also have not much to say. The interesting one is the weak Drive PX2 Autocruise.
First of all, a very eye-catching data is that its power consumption is only 10W. This may be because it uses the Tegra series processor developed by NVIDIA for mobile platforms. However, this power consumption can be said to be very bright anyway, because The level of power consumption of a graded FPGA product is basically at this level, and even higher. And its functional orientation focuses on auto-cruise on the highway, not the entire auto-driving field. It is also well understood that the architecture and extremely low power consumption of a single mobile processor cannot necessarily support ultra-large-scale computing. However, NVIDIA's meaning is not in this.
NVIDIA said that this product can work with NVIDIA's DGX-1. In the picture described by NVIDIA, data scientists can use the NVDIA DGX-1 to train deep neural networks in the data center, and then run the trained neural network on the NVIDIA DRIVE PX2 equipped vehicle. NVIDIA developed exactly the same NVIDIA DRIVEWorks algorithms, libraries, and tools for DGX-1 and Drive PX2 for autocruise. Make it easy for researchers to carry out universal development.
In other words, this computer (Drive PX2 for autocruise) was not designed for development from the beginning, but an application-level device. In other words, NVIDIA began a new phase with its product's actions to include a complete AI development process.
summary
At today's meeting, Huang Renxun emphasized that the P4 is designed to speed up 1U OCP servers with a power of only 50W. The P40 is designed for maximum throughput and can accelerate CPU reasoning by 40 times. Together with Tesla p100 and P4/P40, the ultra-large-scale data center accelerators to be released this year in the United States, the GTC will bring an end-to-end deep learning platform to the data center at both ends of deep learning training and reasoning .
At this point, NVIDIA has formed a training system based on Tesla P100 and DGX-1 in the field of artificial intelligence; a data center inference system with P4/P40 and Tensor-RT as the core; and a smart core with DRIVE PX 2 and Driveworks as the core Driving system. The end-to-end deep learning platform was built through comprehensive layout.
In this conference, the strategic position of the Drive PX2 autocruise may even be higher than P4 and P40. Although it is a self-driving chip, NVIDIA's real intention is to jump out of the AI ​​R&D field as a traditional strength, and prepare and experiment with this product for further development at the AI ​​application level. NVIDIA has clearly become a leader in AI. To achieve this goal, computers designed for specific scenarios like Drive PX2 for autocruise are essential. This product can also be regarded as a test of NVIDIA's water, if it is successful enough, we may soon be able to see NVIDIA in other Internet of Things products.
Three Burner with Built in Hob
3 Burner Gas Cooktop,3 Burner Gas Stove,Automatic Gas Stove 3 Burner,70Cm Gas Cooker Hob
xunda science&technology group co.ltd , https://www.gasstove.be