The essence of artificial intelligence is multi-source big data model training, enabling computers to have technology integration that is partially or beyond human capabilities.
In the computer world, all information can be represented by data, including voice data, text data, image data, pressure data, temperature data, etc. Artificial intelligence is to domesticate computers through massive data to achieve computer imitation, with or beyond human Ability.
Therefore, the collection, cleaning, storage, labeling, processing, and dissemination of massive data have become the key. The development of artificial intelligence requires high-level algorithms and computing power as support, which is reflected in the strong dependence on high-end talents, chips, and data transmission networks.
Algorithms, computing power, and data are the three requirements of artificial intelligence:
-
The base layer provides computing power support, that is, the hardware part;
-
The technical layer provides a general technical platform for algorithm development and domestication of massive data, that is, the software part;
-
The application layer reflects the value of big data domestication in different scenarios.
Driven by data, computing power, and algorithms, technologies related to the global artificial intelligence industry have achieved rapid development, and downstream applications have been continuously enriched.
Among them, computing power represents the ability to process data, is a measure of the sustainable development of digital technology, and is also the core productivity in the era of digital economy.
Next, take the construction of 1000P computing power in the artificial intelligence computing center as an example to deeply understand the computing power of artificial intelligence.
How to understand 1000P?
P is an order of magnitude, 10 to the 15th power, 1000P is 10 billion, and 1000PFlops computing power means 10 billion floating-point calculations per second.
AI computing power unit: magnitude unit + number of calculations per second + data type
It is used to describe the data type involved in the operation, and also expresses the data precision:
-
INT8 8-bit integer data is often suitable for inference operations of deep learning models
-
FP16 16-bit floating-point data is often suitable for training operations of deep learning models
-
FP32 32-bit floating-point data is mainly used in high-performance computing (such as scientific computing)
An AI computing power cluster composed of Atlas 800, taking 1000P computing power as an example, can perform "10 billion trillion" calculations in one clock cycle.
How strong is the computing power of 1000P?
-
Equivalent to 500,000 PCs
-
Taking the data exploration of 200,000 stars as an example, the traditional method takes an experienced scientist 169 days to complete, but now it only takes 10.02 seconds
-
In 26.9 seconds, learn 12 million photos to form a model for image recognition
General-purpose computing and AI computing have different "division of labor" and jointly build diverse computing
| General Computing
The computing power is provided by the CPU. Suitable for complex logic operations, such as most general-purpose software. More than 70% of transistors are used to build Cache and control units, and the number of computing cores ranges from a few to dozens.
General application: office, database, numerical calculation (meteorological forecast, fluid simulation, electromagnetic simulation), etc.
| AI computing
Provide computing power with GPU or NPU. It is suitable for simple logic, calculation-intensive and high concurrent tasks. More than 70% of transistors are used to build computing units, with thousands or tens of thousands of computing cores.
Specific applications: image recognition (face recognition, license plate recognition, action recognition, object detection, perimeter detection, etc.), natural language processing (machine translation, speech recognition, text generation, sentiment analysis, etc.), search recommendation, assisted driving, trends forecast etc.
GPU: Graphics Processing Unit, image processing unit, mainly used for image acceleration and rendering
NPU: Neural Network Processing Unit, Neural Network Processing Unit
Artificial intelligence 1000P is not "1000P" in the sense of supercomputing
Taking high-performance computing as an example, on the TOP500 list, the shortlisted supercomputers must undergo a test called LINPACK to examine the platform's double-precision floating-point computing capabilities.
Taking artificial intelligence as an example, the test program run by artificial intelligence is called Resnet-50, and its results are obtained based on the half-precision floating-point environment, and only the half-precision capability of the platform is examined.
Summary: Artificial intelligence computing is just a branch of high-performance computing. It is aimed at the application requirements of specific fields and specific scenarios for neural networks, deep learning, etc. High-performance computing is the foundation of scientific research applications. It can "calculate heaven, earth and people". Almost all applications can be realized through high-performance computing. The difference between double precision and half precision, double precision is 64 bits, single precision is 32 bits, and half precision is 16 bits.
References for this article: "2022 China Artificial Intelligence Industry Report", "White Paper on the Development of New Computing Power Infrastructure under the East and West Computing", "Interesting Artificial Intelligence Computing Power"
Reprinted: Baidu Security Verification