First Look

Moore’s Law

• The number of transistors on a chip doubles every 18 to 24 months

Types of Computer Systems

• Embedded Systems: Cameras, Cars, Phones, etc.
• Desktop Systems: Personal Computers
• Server Systems: Web Servers, Database Servers, etc.

Seven Great Ideas in Computer Architecture

1. Use abstractions to simplify design

   • Example: Using high-level programming languages like C++ or Java to abstract away low-level details and simplify software development.

2. Make the common case fast

   • Example: Optimizing algorithms and data structures for the most frequently performed operations to improve overall performance.

3. Performance via parallelism

   • Example: Utilizing multi-threading or distributed computing to execute multiple tasks simultaneously and achieve faster execution times.

4. Performance via pipelining

   • Example: Breaking down complex tasks into smaller stages and processing them concurrently to improve throughput and reduce latency.

5. Performance via prediction

   • Example: Implementing branch prediction techniques to minimize the impact of conditional branches on the execution pipeline.

6. Hierarchy of memories

   • Example: Utilizing different levels of cache memory (L1, L2, L3) to store frequently accessed data closer to the processor for faster access.

7. Dependability via redundancy

   • Example: Implementing redundant components or backup systems to ensure system reliability and fault tolerance.

程式處理階層 (Software Hierarchy)

• 應用程式(Application software)：為了達成特定需求(如遊戲、圖片\影片編輯……等等)，由高階語言編成的軟體。
• 系統軟體(System software) ：負責管理及協調硬體階層，包括了作業系統及編譯器等等。
• 硬體層面(Hardware)：囊括了處理器、記憶體，以及IO控制器等等.

程式語言處理階層 (Language Hierarchy)

• 高階語言(High level language)：貼近人類語言，容易閱讀及編寫，透過編譯器或直譯器轉換為機器碼。
• 組合語言(Assembly language)：屬於低階語言，指令多與處理器直接對應，以文字來描述機器的行為，透過組譯器轉換為機械碼。
• 機器碼(Machine code)：由0和1構成的處理器的指令集，電腦可以直接識別。

電腦的元件：

不論是嵌入式、伺服器，或是桌上型電腦，都包含了這5個主要部分：

1. 輸入(Input)
2. 輸出(Output)
3. 記憶體(Memory)
4. 資料路徑(Datapath)
5. 控制(Control)

其中，資料路徑(Datapath)與控制(Control)有時會合併成為處理器(Processor)。

Evaulating Computer Performance

Response Time vs. Throughput

• Response Time(反應時間): Also known as elapsed time, it represents the time required to complete a task. It includes CPU time, I/O, OS overhead, and idle time.
• Throughput(生產能力): The amount of work that can be completed within a certain period of time.

CPU time

$$CPU time = IC \times CPI \times t$$

• IC: Instruction count
• CPI: Cycles per instruction
• t: Clock cycle time

CPI

CPI (Cycle per instruction) ，表示每執行一個指令(Instruction)，所要花費的時脈周期。
Clock cycle time (t) ，表示每一個時脈周期所花費的時間。
Clock Cycles (ClockCycles) ，表示執行所有指令所花費的時脈周期個數。 $$ClockCycles = \sum IC_i \times CPI_i$$
Average CPI: $$avgCPI = \frac{ClockCycles}{IC}$$
指令的個數由編譯器、ISA(ISA instruction set architecture)、程式與演算法共同決定。

CPU Power

$$Power = Capacitance \times Voltage^2 \times Frequency$$

Power Wall: To achieve high frequency and low power, it is necessary to mitigate the impact of both factors. However, after 2004, the techniques of reducing voltage and cooling reached saturation and stagnation. This resulted in a power wall, which posed a significant challenge in improving processor performance. To overcome this challenge, the focus shifted towards the development of multi-core processors.

Amdahl’s Law (阿姆達爾定律)

$$Speedup = \frac{1}{(1 - P) + \frac{P}{S}}$$

• P: Fraction of the program that can be parallelized
• S: Speedup of the parallelizable portion

Since the parallelizable portion of the program is limited by the sequential portion, the speedup is also limited. This is known as Amdahl’s Law.

Cost per die(晶片成本)

• Cost per die:
  $$CostPerDie = \frac{CostPerWafer}{DiePerWafer \times Yield}$$

• Die per wafer: $$DiePerWafer \approx \frac{WaferArea}{DieArea} \approx \frac{\pi \times (WaferDiameter/2)^2}{DieArea}$$

• Yield: $$Yield = \frac{1}{(1-(DefectsPerArea \times DieArea))^N}$$

  • DefectsPerArea: The number of defects per unit area