First, here is the table of contents for the chip learning series.
1. Basic Concepts
-
Basic Classification of Chips: CPU, GPU, FPGA, ASIC, memory chips, analog chips, power chips, etc., their roles and differences in the industry.
-
Design vs Manufacturing: Differences between Fabless, Foundry, and IDM models.
-
Key Process Indicators: Process (nm), yield, power consumption, performance, cost. Why TSMC/Samsung/Intel are positioned in advanced processes.
2. Industry Chain
-
Upstream: EDA tools, IP cores, materials (wafers, photoresist, gases).
-
Midstream: Design companies (e.g., NVIDIA), foundries (TSMC), packaging and testing.
-
Downstream: Application scenarios (mobile phones, automobiles, AI computing power, IoT).
-
Domestic and Overseas Differences: Which links are China’s weaknesses (EDA, lithography machines, advanced processes), and which have breakthroughs (power semiconductors, memory, packaging and testing).
3. Business Models
-
Sources of Profit: One-time sales vs licensing fees (IP/EDA) vs long-term services (SaaS + chip integration).
-
Capital Expenditure: Foundries are extremely asset-heavy, while Fabless is relatively light.
-
Customer Dependence: High concentration in downstream (Apple, Huawei, Tesla, etc.), bargaining power determines the profit stability of chip manufacturers.
-
Cyclicality: The industry is greatly affected by inventory and consumer electronics cycles.
——————————————This article focuses onfirst clarifying the major classifications of integrated circuits (divided by signal processing methods and application characteristics)(1) Digital circuit chips;process discrete logic signals of “0/1”(2) Analog circuit chips;process continuous analog signals (voltage, current, sound, temperature)(3) Mixed-signal chips;involve both analog and digital signal processing (e.g., communication chips, RF chips)For example, in RF chips,the RF front-end modulefirst amplifies weak signals (continuous waves), selects frequency, down-converts, and then entersthe ADC (Analog-to-Digital Converter), converting to a 0/1 digital sequence(4) Special circuit chips.(Military-grade, automotive-grade, radiation-resistant chips, high-temperature/high-pressure environment chips)Among them, logic circuit chips,【General-purpose processor chips】,【memory chips】, system-on-chip (SoC) chips, microcontroller (MCU) chips,【Application Specific Integrated Circuit (ASIC) chips】,【programmable logic devices】all belong todigital circuit chips1. General-purpose processor chips include: CPU, GPU, DSP, etc.,each with its characteristics:1) CPU:Central Processing Unit, the most versatile processor, capable of various calculations, but not the fastest for every taskused to manage, schedule, and control the efficient operation of various components of electronic products and information systems.Typically has only a few cores, with limited parallel capability,2) GPU:Graphics Processing Unit,single-core performance is weaker than CPU, but has superior large-scale matrix computation capabilities compared to CPU,making it suitable for matrix operations and vector calculations, thus widely used in deep learning training and inference (matrix multiplication, convolution operations).Many cores, suitable for parallel processing;3) DPU: Data Processing Unit, generally refers to deep learning processorswhich are more customized than GPUs, with built-in specialized operators (such as matrix multiplication, convolution, attention mechanisms), and lower power consumption: dedicated accelerators designed for AI inference, cloud computing, and network data flow processing.2. Application Specific Integrated Circuit (ASIC) chips are non-general-purpose chips, customized chips. e.g., ID card chips.3. Programmable Logic Devicesinclude PLD, PLA, PAL, GAL,FPGALet’s mainly discuss FPGA (Field Programmable Gate Array) and the main differences with ASIC:1) FPGAcan be reconfigured, circuit logic can be rewritten, sacrificing some efficiency for flexibility;2) ASIClogic is hardwired into the circuit, cannot be modified, and correspondingly, ASIC’s power consumption, speed, and area can be optimized.FPGA comes with ① some programmable logic units + ② programmable interconnectsASIC: AND gates, OR gates, flip-flops, etc., are directly laid out,non-modifiable4. Memory chipsmainly include SRAM, DRAM, LPDDR, PROM, flash memory chips, etc.Differences lie in: speed, power consumption, volatility.1) SRAM:Static Random Access Memory, stores data using flip-flops
-
Advantages: Fast speed, low latency.
-
Disadvantages: High cost, small capacity, high power consumption.
Commonly used asCPU/GPU cache (Cache)2) DRAM:stores data using capacitors, charge leaks → must be refreshed constantly.
-
Advantages: Large capacity, low cost.
-
Disadvantages: Slower than SRAM, and requires refreshing.
Commonly used: Main memory (memory sticks) in computers and servers.
3) LPDDR (Low Power DDR)is a mobile-optimized version of DRAM, significantly reducing power consumption.
-
Advantages: Low power consumption, suitable for battery-operated devices.
-
Disadvantages: Slightly slower than high-end desktop DDR, but with significantly reduced energy consumption.
Commonly used: Smartphones, tablets, automotive systems.
4) PROM (Programmable ROM)is blank at the factory, users can write data once, and it cannot be changed afterwards.
-
Advantages: Data is reliably stored and not lost.
-
Disadvantages: Can only be written once, cannot be modified.
Commonly used: Firmware, initialization programs at the time of device manufacturing.
5) Flash Memory
Non-volatile memory (data is not lost when power is off), based on charge storage.
-
Advantages: Large capacity, can be repeatedly erased and written, moderate cost.
-
Disadvantages: Limited write lifespan, slower than DRAM.
Commonly used: USB drives, SSDs, mobile storage, camera memory cards.
In robotics and artificial intelligence scenarios, what specific chips are needed?
1. Computing Power Chips
GPU: Processes visual, speech recognition, and other AI inference/training;
DPU / NPU: Deep Learning Processing Unit / Neural Processing Unit, designed for deep learning inference
2. Analog Chips (ADC/DAC): Converts analog signals from cameras, microphones, and tactile sensors into digital signals.
Sensor Chips: IMU (Inertial Measurement Unit), gyroscopes, accelerometers, pressure sensors.
Mixed-Signal Chips: Processes radar and LiDAR signals.
3. Power Chips
Power Semiconductors: Control motors, batteries, and power management
MCU (Microcontroller): Each joint/actuator requires an independent MCU for real-time control (low latency, low power consumption)
4. Memory Chips
SRAM / DRAM: Used as high-speed cache and memory during AI inference.
Flash: Stores the robot operating system and trained model parameters.
5. CommunicationFPGA: Used for high-speed data interfacesRF chips, etc.Above is an introduction to the basic situation and applications of various major chips.The next chapter:Design vs Manufacturing: Differences between Fabless, Foundry, and IDM models.—————————— The customary voice-over ——————————
Uh, it was originally very quiet, thinking a bit in the dead of night, got up to study in the middle of the night…