A
| Abbreviation | Full Name | Meaning |
| API | Application Programming Interface | Some predefined functions, or conventions used to connect different components of a software system. |
| ARM | Advanced RISC Machine | The first RISC microprocessor designed by Acorn Ltd. in the UK with low power consumption cost. Its main design goals are low cost, high performance, and low power consumption. It is designed in 32-bit, but also equipped with a 16-bit instruction set, which generally saves up to 35% than the equivalent 32-bit code while retaining all the advantages of 32-bit systems. |
| ASIC | Application Specific Integrated Circuit | It refers to the integrated circuits designed and manufactured in response to specific user requirements and the needs of specific electronic systems. |
B
| Abbreviation | Full Name | Meaning |
| BN | BatchNorm/BatchNorm2d | Batch normalization is a practice of unifying scattered data to improve the performance and stability of artificial neural networks, which is a technique that provides zero mean/unit variance input to any layer in a neural network to optimize the neural network. |
| BPU | Brain Processing Unit | A self-developed programmable acceleration engine of Horizon Robotics, which has iterated on the following computing platform products: Computing platforms: Sunrise I, Sunrise II, Journey II, Sunrise III, Journey III, Journey V and Journey VI. FPGA computing platform: Matrix 1.0, Matrix 1.6, Matrix 2.0, etc. |
C
| Abbreviation | Full Name | Meaning |
| CPU | Central Processing Unit | One of the main devices of a computer, the function is mainly to interpret computer instructions and to process data in computer software. As the computing and control core of the computer, it is the final execution unit for information processing and program operation. |
D
| Abbreviation | Full Name | Meaning |
| DDR | Double Data Rate SDRAM | SDRAM transfers data once in a clock cycle, while DDR transfers data twice in a clock cycle, once on the rising edge and once on the falling edge, a concept known as prefetching. MT/S units are generally used when describing DDR speed, meaning how many megabytes of data transfer per second. |
| DL | Deep Learning | A new research direction in the field of machine learning, deep learning is the process of learning the intrinsic laws and manifestation levels of sample data. The information obtained from these learning processes can be of great help in the interpretation of data such as text, images and sounds. Its ultimate goal is to enable machines to learn analytically, like humans, and to recognize data such as text, images, and sounds. |
| DMA | Direct Memory Access | It is an important feature of all modern computers that it allows hardware devices of different speeds to communicate without relying on the massive interrupt load of the CPU. Otherwise, the CPU needs to copy each slice of information from the source to a temporary storage area and then write them back again to the new place. During this process, the CPU is unable to perform other operations. |
F
| Abbreviation | Full Name | Meaning |
| FPS | Frames Per Second | Frames recognized accurately per second, used to reflect the computation efficiency of a processor. Under the same precision, higher FPS means faster processing speed, and vice versa. Horizon believes that compute power does not fully reflect the real performance of a processor. It is like, for a car, horsepower (in HP) is not as good as 0-100 km acceleration time (in seconds) to reflect the real power performance of the car, for an automotive processor, the compute power (unit: TOPS) does not reflect the actual performance, but the frame per second (unit: FPS) is a more realistic performance indicator. |
H
| Abbreviation | Full Name | Meaning |
| HBDK | Horizon BPU Development Kit | Horizon's application development toolchain for BPU processors, which contains compilers and a series of development tools for convenient heterogeneous development, such as runtime library (Runtime library) for BPU optimization CPU master core compiler in BPU, BPU simulator, and APIs shielding the complex details of heterogeneous hardware. |
| HBM | Horizon BPU Model | Model file generated by HBDK compiler. The HBM contains information such as BPU commands, model parameters, and input and output descriptions. |
| HBIR | Horizon BPU Intermediate Representation | HBDK-defined model intermediate representation (IR). |
| HBRT | Horizon BPU Runtime | The runtime of the HBDK compilation toolset, usually used as one of the underlying base setups for other embedded development toolkits, provides a series of APIs used to enable upper-level code to parse the HBM file, generate the configuration information needed to drive the BPU and process the running results of the BPU model. |
| HMCT | Horizon Model Convert Toolkit | HMCT is a model conversion tool provided by Horizon, which completes the post-quantization of the model through the process of parsing, optimization, and calibration of the model. |
I
| Abbreviation | Full Name | Meaning |
| IDMA | Integrated Direct Memory Access | Integrated DMA controller. |
| IPC | Instructions Per Cycle | In computer architecture, instructions per cycle is an indicator to measure the processor performance, i.e., the average number of instructions executed per clock cycle. IPC is the multiplicative inverse of cycles per instruction (CPI). |
L
| Abbreviation | Full Name | Meaning |
| LMDB | Lightning Memory-Mapped Database | LMDB uses memory mapped files to provide better input/output performance. LMDB files can be opened by multiple processes at the same time, with very high data access speed and simple access. There is no need to run a separate database management process, just reference the LMDB library in the code that accesses the data by providing the file path. |
P
| Abbreviation | Full Name | Meaning |
| PTQ | Post-training Quantization | A post-training quantization method that first trains a floating-point model and then uses a calibrated picture to calculate quantization parameters to convert the floating-point model into a quantized model. |
Q
| Abbreviation | Full Name | Meaning |
| QAT | Quantization Aware Training | The floating-point model structure is first intervened during the floating-point training, allowing the model to perceive the loss from the quantization and reduce the accuracy loss brought by the quantization. |
R
| Abbreviation | Full Name | Meaning |
| RPC | Remote Procedure Call | It is a protocol for requesting services from the remote computer programs over the network without the need to understand the underlying network technologies. |
| RPMSG | Remote Processor Messaging | A virtio-based message passing BUS, which defines a standard binary interface, used for core-to-core communication in heterogeneous multicore processing systems, allowing kernel drivers to communicate with remote processors available on the system. |
S
| Abbreviation | Full Name | Meaning |
| SDK | Software Development Kit | In general, it is a collection of development tools that are used by software engineers to build application software for specific software packages, software frameworks, hardware platforms, operating systems, etc. |
| SOTA | state-of-the-art | The model for achieving the current optimal result on a task in machine learning. For example, in an image classification task, if a model achieves the current optimal performance on a commonly used dataset (e.g., ImageNet), we can say that the model has achieved SOTA. |