1 Basic concepts
1.1 Computing power unit
TOPS: Refers to the number of integer operations that can be performed per second. It represents the computer's processing power when processing images, audio and other tasks. The unit of TOPS is trillion operations per second. Generally refers to the integer arithmetic abilityINT8.
TFLOPS: Refers to the number of floating-point operations that can be performed per second. It represents the computer's processing power when processing tasks such as scientific computing and machine learning. The unit of TFLOPS is trillion floating point operations per second. Generally refers to single precision performanceFP32.
MFLOPS: Equivalent to one million floating point operations per second. (megaFLOPS)
GFLOPS: Equivalent to one billion floating point operations per second. (gigaFLOPS)
PFLOPS: Equivalent to one quadrillion floating-point operations per second. (petaFLOPS)
1.2 Memory bandwidth
Memory bandwidth refers to the speed at which data can be transferred from the computer's motherboard to memory. It is often used as one of the measures of memory performance to determine how quickly a computer can read and write data into memory.
The size of the memory bandwidth depends on the product of the memory bus width and the memory frequency. For example, if the memory bus width is 64 bits and the memory frequency is 1600 MHz, the memory bandwidth is 64 bits x 1600 MHz = 12.8 GB/s.
The amount of memory bandwidth directly affects a computer's computing power because it can limit the amount of data the computer can read and write quickly. If there is insufficient memory bandwidth, the computer may have to wait longer to read or write data, making the computer less efficient. Especially when working with large data sets and running memory-hungry applications, the amount of memory bandwidth becomes even more important. Therefore, when choosing a computer configuration, you need to consider memory bandwidth and other key metrics to ensure that the computer can meet your specific computing needs.
1.3 DMIPS
DMIPS (Dhrystone Million Instructions Per Second, millions of machine language instructions processed per second) describes the computing power of the CPU.
The filtering algorithm of multi-sensor fusion for autonomous driving, the registration algorithm of laser point cloud, and most path planning and decision-making algorithms test the computing power of the CPU.
1.4 ASIC
ASIC (Application-Specific Integrated Circuit) refers to an application-specific integrated circuit, which is a customized integrated circuit for specific applications. Unlike general-purpose integrated circuits (such as microprocessors), ASICs are designed and manufactured for specific applications, and the performance and power consumption of the circuit can be optimized to meet the requirements of the specific application. ASICs are commonly used in high-performance computing, communications, graphics processing and other fields.
ASIC design requires multiple steps such as circuit design, logic design, physical design, and verification, so the manufacturing cost is high, but it can achieve high performance and reliability in high-end applications. Compared with FPGA, ASIC has advantages in performance, power consumption and area, but lacks the flexibility and programmability of FPGA.
ASIC accelerators include CVP (Computer Vision Processors), DLA and MA (Multithreaded Accelerator), of which 18 CVPs are Mobileye's next-generation ASIC chips designed for traditional computer vision algorithms. The overall AI computing power of the chip is designed to be only 24TOPS.
For example: DLA deep learning acceleration in Jetson is an ASIC.
2 GeForce 30x series
3 Jetson
Reference 1: AI chip, the "Water Margin card" in autonomous driving
Reference 2: Core parameters