Multidimensional information processing device and general-purpose artificial intelligence

Abstract

We provide information processing devices that achieve high speed, low power consumption, low heat generation, compact size, and low cost for stored program computers. [Solution] The multidimensional information processing device 200 has a multidimensional processing information monitoring device DPU0 neuron and neurons & synapses DPU1 to n that store multidimensional processing decision table language programs, and the neurons & synapses have a condition stub CS that stores conditional statements and calculation formulas, a condition entry CE that stores YES, NO, and calculation values, an action stub AS that stores execution commands, and an action entry AE that stores execution order, the multidimensional processing information monitoring device has a flag that manages whether the neurons & synapses are operating or dormant, and the neurons & synapses store multiple multidimensional processing decision table language programs, variable items, and fixed value items, and calculations are performed in memory by the neurons & synapses.

Description

[Technical Field]

[0001] The current 1.5-dimensional programming language for computers, influenced by 1-dimensional voice, 2-dimensional murals, paper, blackboards, electronic device screens, etc., By expressing it in a multidimensional decision table language, This is a technical field related to increasing the speed, power consumption, heat generation (eliminating water resource shortages), miniaturization, and cost of stored-program computers such as semiconductor computers and quantum computers. [Background technology]

[0002] (Regarding the multidimensional processing device filed on December 29, 1979) Mathematical equations and speech are one-dimensional, while paper, blackboards, and computer screens are two-dimensional. Therefore, it was a 1.5-dimensional programming language for current computers, He invented a multidimensional processing decision table language, and in an era before the Internet, he created a high-speed information processing device with router-like functions and multidimensional (3D configuration) operations. Rather than inputting and recording a 1.5-dimensional programming language written on two-dimensional paper, the operation involves inputting and recording a multidimensional processing decision table language from an information device screen (now a PC screen) into a multidimensional processing device (a three-dimensional computer) and performing calculations. The present invention takes into consideration its compatibility with current supercomputers and artificial intelligence (AI).

[0003] (Regarding processing speed) Mathematical equations are said to have been found in rock paintings on cave walls from the Mesopotamian civilization, dating back about 6,000 years ago. There are various ways to write mathematical formulas, such as on a sheet of paper, on a blackboard, or on an electronic device screen. Naturally, everything is two-dimensional (surfaces), and mathematical formulas are based on one-dimensional formulas (lines), and are expressed and proven in 1.5 dimensions using multiple lines. Both Einstein and von Neumann thought about and expressed mathematical formulas in 1 and 1.5 dimensions, which is why stored-program computers cannot execute in 3 or multiple dimensions. In the brains of humans and animals, neurons and synapses work together to process information in a multidimensional (at least three-dimensional) manner. Current computers, known as von Neumann computers, use programming languages ​​based on 1.5-dimensional mathematical formulas, with the central processing unit (CPU) executing each line sequentially. In today's supercomputers, such as artificial intelligence (AI), some programs use multiple central processing units (CPUs) and memory storage devices to perform distributed simultaneous processing of the mathematical formula program calculations that can be processed in parallel. In other words, when executing a program for a general-purpose system, the central processing unit CPU and the storage device memory cannot be efficiently distributed for processing. The speed at which electricity travels through an ideal conductor with no resistance is approximately 300,000 km / s, the same speed as light. The nerve conduction speed in the human brain is known to be approximately 0.2 to 1.5 m / s. Although the nerve conduction speed of the human brain is slower than that of the central processing units (CPUs) of current computers, because information is processed multidimensionally (at least in three dimensions), it feels as fast as a semiconductor computer, and it operates on low energy, which is an important hint for technological improvement. When executing mathematical formulas that can be processed in parallel for artificial intelligence (AI), if all lines are executed simultaneously, a speed increase of tens of thousands of times can be expected. In the case of general-purpose systems, however, execution is performed sequentially, one line at a time, which is slow. However, even semiconductor computers can process information faster by operating in a multidimensional processing decision table language. Existing computers have a central processing unit (CPU) that stores approximately one program and tens of thousands of lines of data in a storage device such as memory, and when executed, copies each line to the CPU and the data items, which are then executed. This makes them slow. The human brain is three-dimensional, and when we first experience something, there is a stage where it is stored in the human brain, but the data items (eye images, ear sounds, various senses) are moved and copied to the pre-stored, processing program-like cellular tissue neurons and synapses, and information is processed three-dimensionally. Therefore, it is fast. In the age of wired logic computers, the programming language portion was stored in advance, and only variable items were input as data. In other words, the program is written in a multidimensional processing decision table language and the fixed value data items are moved and copied, and the program is recorded and stored in the CPU and memory in advance. Information processing is required. To the general-purpose artificial intelligence (AGI), we think with the human brain (3D), Communicating using natural language (1-dimensional) or existing programming languages ​​(1.5-dimensional) is insufficient for both the human inputter and the AGI receiver, and a new multidimensional processing programming language is desired. This can be applied to protein computers, which have been studied for a long time and are expected to be the next step after vacuum tube computers, semiconductor computers, superconducting quantum computers, and optical quantum computers.

[0004] (Low power consumption, low heat generation (eliminating water resource shortages), compact size, and low cost) von Neumann semiconductor computers are capable of high-speed, accurate calculations and consume large amounts of power, but currently only support 1.5-dimensional programming languages. The human brain neuron system has a slow nerve conduction speed, but is three-dimensional, capable of high-speed calculations, inaccurate calculations, low power consumption, and compact size. Quantum computers perform ultra-fast but inaccurate calculations and consume enormous amounts of power, but they currently have a 1.5-dimensional structure. The human brain operates on about 20 watts, while a data center doing the same job consumes megawatts of power. The big bang of artificial intelligence (AI) began in 2012 when Professor Hinton of the University of Toronto developed a deep learning method that is equivalent to the processing in the human brain. However, the human brain processes information quietly using limited energy, and it is not like we sweat or develop a high fever just because we have thought things through deeply. For an artificial intelligence (AI) engine to truly compete with the processing power of the human brain, there are naturally limits to how far it can go in trying to win by brute force using large amounts of high-speed semiconductors and huge amounts of electricity, as is currently the case. Artificial intelligence (AI) consumes a huge amount of electricity, and while it cannot be said that global warming is the sole cause, there are concerns that it will be depleted due to massive typhoons, tornadoes, forest fires, and changes in the habitats of plants and animals. In addition, the memory calculation technology "Palm-sized AI data centers" is being developed by former Hitachi engineers, who claim that the memory calculation technology will reduce the power consumption of existing systems to one-thousandth and the cost to one-tenth. It is claimed that it will be possible to miniaturize AI data centers to the size of the palm of your hand. Flodia's CiM technology is characterized by consuming approximately one-thousandth of the power consumption of NVIDIA's GPU (graphics processing unit). While GPUs are mainstream for artificial intelligence (AI), a major issue is that the power consumed by memory data transfer accounts for a large portion of the total power consumption. CiM performs deep learning within the memory, which can dramatically reduce power consumption. The company aims to achieve this by using its own proprietary flash memory, which is resistant to charge leakage. The current trend in artificial intelligence (AI) is compactness. Examples include "quantization," which reduces the number of bits that indicate the accuracy of learning and inference, and China's DeepSeek, which has achieved LLM (large-scale language model) at low cost. A similar trend is emerging in hardware. Flodia's vision is to implement LLM through 3D integration of CiM chips. If this is realized, it will be possible to run the US OpenAI LLM "GPT-4" without a huge data center, and it will be small enough to fit in the palm of your hand. The advantages of CiM are its low power consumption and low cost. Floadia's CiM technology not only consumes approximately 1 / 1000th of the power of a GPU, but also has manufacturing costs that are 1 / 10th of those of digital computing methods.

[0005] (Regarding accurate information processing) Furthermore, like current quantum computers, the human brain has unstable information processing and frequently produces answers containing errors. To address this issue, the human brain sends the same information from each neuron and synapse to approximately eight (2 to the power of 3) neighboring neurons and synapses, which then process the information. It is believed that the correct answer is determined by majority vote and processed as the correct answer. However, when multiple pieces of information are processed simultaneously, it can sometimes become difficult to distinguish between them. In superconducting quantum computers, there is a method that solves the instability of information processing by treating 64 physical qubits as one logical qubit. In other words, quantum computers are not suitable for systems that support conventional social infrastructure. Although the speed of quantum computers is attractive, they cannot perform calculations as accurately as semiconductor computers. The multidimensional information processing device made of the semiconductor of the present invention does not suffer from instability in information processing, and does not require the above-mentioned measures for stabilizing information processing. [Prior art documents] [Patent documents]

[0006] [Patent Document 1] Patent Publication No. 58-53778 (Title of invention: Multidimensional processing device) Summary of the Invention [Problem to be solved by the invention]

[0007] The challenges of increasing the speed, power consumption, heat generation (resolving water resource shortages), miniaturization, and cost of semiconductor and stored-program computers such as quantum computers are becoming more and more pressing, and there is a demand for solutions that are compatible with existing computers. Artificial intelligence (AI) consumes a huge amount of electricity, and while it cannot be said that global warming is the sole cause, there are concerns that it will be depleted due to massive typhoons, tornadoes, forest fires, and changes in the habitats of plants and animals. It is hoped that these inventions, which have been made available free of charge, will provide hints for new implementation technologies. [Means for solving the problem]

[0008] It is compatible with the architecture of existing semiconductor computers and quantum computers. By providing a multidimensional (3D) information processing device using the multidimensional processing decision table language applied in "Patent Publication No. 58-53778 (Multidimensional Processing Device) filed on December 29, 1979," it is possible to perform distributed processing more efficiently and generally than supercomputers such as artificial intelligence (AI) with current architecture, and to perform high-speed parallel processing with low power consumption, low heat generation (eliminating water resource shortages), compact size, and low cost. [Effects of the Invention]

[0009] It is compatible with existing semiconductor computers and quantum computers. By providing a multidimensional information processing device using a multidimensional processing decision table language, Rather than supercomputers such as artificial intelligence (AI) with current architecture, It enables general-purpose, efficient distributed processing, low power consumption, low heat generation (eliminating water resource shortages), compact size, low cost, and high-speed parallel processing. High-speed multidimensional information processing enables accurate and rapid tsunami forecasts immediately after a major earthquake. Artificial intelligence (AI) consumes a huge amount of electricity, and while it is not the sole cause of global warming, it can prevent depletion caused by massive typhoons, tornadoes, forest fires, and changes in the habitats of plants and animals. Furthermore, the multidimensional processing decision table language allows mathematical expressions from higher to lower concepts to be written and executed in a multidimensional manner. In addition, maintenance work will be possible by automatically converting old development languages ​​of currently operating computers into multidimensional processing decision table languages ​​using artificial intelligence (AI), and then automatically converting them into new development languages ​​of recent years using existing technologies (script language technology and variable determination methods (Patent Publication No. 2002-312167)). [Brief explanation of the drawings]

[0010] [Figure 1] Diagram of the multidimensional processing decision table language [Figure 2] Conceptual diagram of an example implementation of a multidimensional information processing device [Figure 3] Diagram of the multidimensional processing device (Patent Publication No. 58-53778) (reference) DETAILED DESCRIPTION OF THE INVENTION [Example] [Example]

[0011] It is applied to semiconductor computers, superconducting quantum computers, and optical quantum computers, which are faster, consume less power, generate less heat (eliminates water resource shortages), are smaller, and are lower in cost. (Regarding speed improvement) In supercomputers and artificial intelligence (AI) data centers, semiconductor chips are mounted on silicon substrates, glass substrates, and resin organic substrates using the following three advanced methods. (1) Silicon interposer type, (2) Organic interposer type, (3) Silicon bridge type, using advanced 2D mounting. Advanced 2D packaging is also known as 2.5D packaging or 2.5D packaging. Alternatively, as a method that exists and is currently in operation in existing products, (1) multiple semiconductor chips are mounted on a motherboard as in the past, (2) connected to each node of a supercomputer / artificial intelligence (AI) data center, and (3) connected via an intranet / internet. By implementing the above-mentioned advanced technologies and existing product technologies, It is composed of a central processing unit (CPU) and memory as an information monitoring device. Multidimensional processing information monitoring device DPU0 neuron (201), It consists of a central processing unit (CPU), a graphics processing unit (GPU), a general-purpose processing unit (GPGPU), and memory. DPU1 to 8(n) neurons and synapses (202 to 208(n)) that store multidimensional processing decision table language programs are arranged. The DPU1 to 8(n) neurons and synapses (202 to 208(n)) that store the multidimensional processing decision table language program described above are Condition stub CS(104) stores conditional statements and calculation formulas, YES · NO · Condition entry CE (105) to store the calculated value, Action stub AS (106) that stores execution instructions (calculation formulas, branch instructions, and the number of program lines for estimating the processing time for each DPU free time prediction), It has an action entry AE (107) that stores the execution order, For example, it stores tens of thousands of lines of programming language description, The multidimensional processing information monitoring device DPU0 neuron (201) as the information monitoring device has a flag that manages whether the DPU1 to 8(n) storing the multidimensional processing decision table language program is operating or inactive, The multidimensional processing information monitoring device DPU0 neuron (201) as the information monitoring device includes: The DPUs 1 to 8(n) storing the multidimensional processing decision table language program each have a flag for managing whether the DPU is in operation or inactive, The DPU1 to 8(n) neurons and synapses (202 to 208(n)) that store the multidimensional processing decision table language program described above are storing a plurality of multidimensional processing decision table language programs and variable items and fixed value items; This is an information processing device characterized by in-memory calculations by each DPU (CPU, GPU, GPGPU and memory) of DPU1 to 8(n) neurons & synapses (202 to 208(n)) that store the multidimensional processing decision table language program. Multidimensional processing decision table languages ​​include: Condition stub CS (conditional statements, calculation expressions (integer values ​​and floating-point operations)), Condition entry CE (YES / NO / calculated value (floating point error allowed)), Action stub AS (execution command (calculation formula, branch destination DPU or IP address, number of program processing lines for rough estimation of processing time for DPU free time prediction)), There is an entry field for Action Entry AE (execution order entry). It is important to note that variable data items are divided into integer calculation data items (abacus-like, no calculation error) and floating-point calculation data items (slide rule-like, high-speed calculation, but calculation error occurs). By providing a multidimensional processing device using a mathematical formula decision table multidimensional language for general-purpose computers, we can achieve general-purpose, efficient distributed processing and high-speed parallel processing, more than supercomputers such as artificial intelligence (AI) with current architectures. In addition, a multidimensional processing decision table language is used to describe and execute mathematical expressions from higher to lower concepts in a multidimensional manner. (Low power consumption, low heat generation (eliminating water resource shortages), compact size, and low cost) In a semiconductor computer that is not information processing, which is made up of a huge number of proteins in the human brain, if there are not enough DPU1 to 8(n) neurons and synapses (202 to 208(n)) that store multidimensional processing decision table language programs, By storing a plurality of multidimensional processing decision table language programs and registering and arranging each operation command and fixed data item at the address where each multidimensional processing decision table language is stored, DPU1-8(n) neurons and synapses (202-208(n)) that store other multidimensional processing decision table language programs are active or inactive. The multidimensional processing information monitoring device DPU0 neuron (201) as an information monitoring device makes a decision, searches for DPU1 to 8(n) that have stored a dormant multidimensional processing decision table language program, executes multidimensional processing, performs high-speed calculations such as tens of thousands of times, and achieves low power consumption, low heat generation (eliminating water resource shortages), compact size, and low cost. In order to achieve the same state as in the era of wired logic computers, Multidimensional processing information monitoring device DPU0 neuron (201), The multidimensional processing decision table language program is stored in advance in DPU1 to DPU8(n) neurons and synapses (202 to 208(n)) during the first loop. (Regarding development methods) Automatically convert existing programming languages ​​of currently operating computers into multidimensional processing decision table languages, or into new development languages ​​of recent years. To facilitate maintenance work. The programming language written in each stub of the multidimensional processing decision table language is COBOL, Java, and other languages ​​are converted to, for example, Dart and pre-compiled. The Dart language allows the same objects to run on various operating systems (Windows, Android, iOS, etc.). We are now in an age where AI such as Google's Gemini can quickly develop a multidimensional processing decision table language (design), its compiler software, an AI prompt editor, and a tool that automatically converts the existing development language of a running computer into a multidimensional processing decision table language, in order to apply multidimensional information processing equipment to existing supercomputers and AI. [Explanation of symbols]

[0012] 101 General von Neumann programming languages 102 Multidimensional Processing Decision Table Language 103 Construction of a Multidimensional Processing Decision Table Language 104 Condition Stab CS 105 Condition Entry CE 106 Action Stab AS 107 Action Entry AE 200 Conceptual diagram of an example implementation of a multidimensional information processing device 201 Multidimensional processing information monitoring device DPU0 neuron 202~208(n) DPUs storing multidimensional processing decision table language programs 1~8(n) neurons & synapses

Claims

1. In supercomputers and artificial intelligence (AI) data centers, Semiconductor chips are mounted on silicon substrates, glass substrates, or resin organic substrates, or As in the past, multiple semiconductor chips were mounted on a motherboard, or connected to each node of a supercomputer or AI data center, or in an implementation environment connected via an intranet or the Internet. It is composed of a central processing unit (CPU) and memory as an information monitoring device. A multidimensional processing information monitoring device DPU0 neuron (201); Central processing unit CPU, image processing unit GPU, general-purpose processing unit GPGPU, configured in memory DPU1 to DPU8(n) neurons and synapses (202 to 208(n)) storing a multidimensional processing decision table language program are arranged. The DPU1 to DPU8(n) neurons and synapses (202 to 208(n)) that store the multidimensional processing decision table language program include: A condition stub CS (104) that stores conditional statements and calculation formulas; A condition entry CE (105) for storing YES, NO, and calculated values; An action stub AS (106) that stores execution instructions (calculation formulas, branch instructions, and the number of program lines for estimating the processing time for each DPU free time prediction); It has an action entry AE (107) that stores the execution order, The multidimensional processing information monitoring device DPU0 neuron (201) as the information monitoring device has a flag that manages whether DPU1 to DPU8 (n) storing the multidimensional processing decision table language program is in operation or inactive, The multidimensional processing information monitoring device DPU0 neuron (201) as the information monitoring device includes: The DPUs 1 to 8(n) storing the multidimensional processing decision table language program each have a flag for managing whether the DPU is in operation or inactive; The DPU1 to DPU8(n) neurons and synapses (202 to 208(n)) that store the multidimensional processing decision table language program include: storing a plurality of multidimensional processing decision table language programs and variable items and fixed value items; The multidimensional processing decision table language program is stored in the DPUs 1 to 8 (n) and the neurons and synapses (202 to 208 (n)) perform in-memory calculations.

1. An information processing device comprising:

Images