CPU基础知识详解

冯·诺依曼计算机

冯·诺依曼计算机由存储器、运算器、输入设备、输出设备和控制器五部分组成。

[图片上传失败...(image-763f0e-1679189920085)]

哈佛结构

哈佛结构是一种将程序指令存储和数据存储分开的存储器结构，它的主要特点是将程序和数据存储在不同的存储空间中，即程序存储器和数据存储器是两个独立的存储器，每个存储器独立编址、独立访问，目的是为了减轻程序运行时的访存瓶颈。哈佛架构的中央处理器典型代表ARM9/10及后续ARMv8的处理器，例如：华为鲲鹏920处理器。

[图片上传失败...(image-618186-1679189920085)]

组成计算机的基础硬件都需要与主板（Motherboard）连接

[图片上传失败...(image-6e335c-1679189920085)]

计算机基础硬件 (2)

[图片上传失败...(image-26452a-1679189920085)]

Opening the Box（Apple IPad2）

[图片上传失败...(image-c996fb-1679189920085)]

手机的内部结构 – 华为Mate30 Pro

主板(来自于 Tech Insights）

[图片上传失败...(image-62a105-1679189920085)]

主板背面

[图片上传失败...(image-7cfe27-1679189920085)]

射频板

[图片上传失败...(image-cff0e7-1679189920085)]

Inside the Processor (CPU)

Datapath(数据通路): performs operationson data
Control: sequences datapath, memory, ...
Register 寄存器
Cache memory 缓存
- Small， fast： SRAM(静态随机访问存储器)
  memory for immediate access to data

[图片上传失败...(image-132d6d-1679189920085)]

Intel Core i7-5960X

[图片上传失败...(image-738d5-1679189920085)]

毅力号CPU曝光：250nm工艺、23年旧架构、主频仅233MHz

毅力号搭载的处理器是20多年前技术的产品。处理器型号为PowerPC 750处理器，与1998年苹果出品的iMac G3 电脑同款，PowerPC 750 处理器最高主频速度仅233MHz，且晶体管数量也只有600 万个，但单价仍高达20 万美元（约130万元）。抗辐射、耐寒冷-55~125℃

对比苹果最近推出的M1ARM 架构处理器拥有最高主频3.2GHz，晶体管数量达160 亿个。

[图片上传失败...(image-411412-1679189920085)]

处理器发展趋势

[图片上传失败...(image-aa8ff1-1679189920085)]

主流CPU发展路径

[图片上传失败...(image-f3b278-1679189920085)]

Through the Looking Glass

LCD screen: picture elements (pixels像素)

Mirrors content of frame buffer memory帧缓冲存储器

[图片上传失败...(image-a9b7c3-1679189920085)]

Touchscreen(触摸屏)

PostPC device

Supersedes(取代)keyboard and mouse
Resistive阻性 and Capacitive容性types
- Most tablets, smart phones use capacitive
- Capacitive allows multiple touches simultaneously(多点同时触控)

[图片上传失败...(image-23f718-1679189920085)]

A Safe Place for Data

Volatile main memory(易失性主存)

Loses instructions and data when power off(断电)

Non-volatile secondary memory

Magnetic disk(磁盘)
Flash memory(闪存)
Optical disk (CDROM, DVD) 光盘

Networks 与其他计算机通信

Communication(通信), resource sharing(资源共享), nonlocal access(远程访问)
Local area network (LAN): Ethernet,局域网/以太网
Wide area network (WAN): the Internet，广域网/互联网
Wireless network: WiFi, Bluetooth(蓝牙)

[图片上传失败...(image-23db-1679189920085)]

计算机基础硬件 (3)
[图片上传失败...(image-ae35e1-1679189920085)]

Abstractions抽象

The BIG Picture

Abstraction helps us deal with complexity
- Hide lower-level detail
Instruction set architecture (ISA)指令集体系结构
- The hardware/software (abstraction) interface
Application< ---- > binary interface应用二进制接口
- The ISA plus system software interface
Implementation(区别于Architecture)
- The details underlying the interface

半导体与集成电路

Technology Trends 处理器和存储器制造技术--趋势

Electronics technology continues to evolve

Increased capacity and performance
Reduced cost

[图片上传失败...(image-c3976c-1679189920085)]

[图片上传失败...(image-6eb5ec-1679189920085)]

Semiconductor Technology

Silicon硅: semiconductor 半导体
Add materials to transform properties属性:
- Conductors
- Insulators
- Switch

[图片上传失败...(image-d21f28-1679189920085)]

[图片上传失败...(image-8b1bb0-1679189920085)]

设备列表

厂商在制造芯片的过程中，从前端工序、到晶圆制造工序，之后再到封装和测试工序，主要用到的设备依次包括，单晶炉、气相外延炉、氧化炉、低压化学气相沉积系统、磁控溅射台、光刻机、刻蚀机、离子注入机、晶片减薄机、晶圆划片机、键合封装设备、测试机、分选机和探针台等

集成电路发明

1952年，英国雷达研究所的科学家达默在一次会议上提出：可以把电子线路中的分立元器件，集中制作在一块半导体晶片上，一小块晶片就是一个完整电路，这样一来，电子线路的体积就可大大缩小，可靠性大幅提高。这就是初期集成电路的构想。
1956年，美国材料科学专家富勒和赖斯发明了半导体生产的扩散工艺，这样就为发明集成电路提供了工艺技术基础。
1958年9月，美国德州仪器公司的青年工程师杰克·基尔比（Jack Kilby），成功地将包括锗晶体管在内的五个元器件集成在一起，基于锗材料制作了一个叫做相移振荡器的简易集成电路，并于1959年2月申请了小型化的电子电路（Miniaturized Electronic Circuit）专利（专利号为No.31838743，批准时间为1964年6月26日），这就是世界上第一块锗集成电路。

[图片上传失败...(image-582dda-1679189920085)]

2000年，集成电路问世42年以后，人们终于了解到他和他的发明的价值，他被授予了诺贝尔物理学奖。诺贝尔奖评审委员会曾经这样评价基尔比：“为现代信息技术奠定了基础”。
1959年7月，美国仙童半导体公司的诺伊斯，研究出一种利用二氧化硅屏蔽的扩散技术和PN结隔离技术，基于硅平面工艺发明了世界上第一块硅集成电路，并申请了基于硅平面工艺的集成电路发明专利（专利号为No.2981877，批准时间为1961年4月26日。虽然诺伊斯申请专利在基尔比之后，但批准在前）。
基尔比和诺伊斯几乎在同一时间分别发明了集成电路，两人均被认为是集成电路的发明者，而诺伊斯发明的硅集成电路更适于商业化生产，使集成电路从此进入商业规模化生产阶段。

Intel Core i7 Wafer

300mm wafer, 280 chips, 32nm technology
Each chip is 20.7 x 10.5 mm

[图片上传失败...(image-e6dac4-1679189920085)]

Integrated Circuit Cost

$Cost per die =\frac{\text { Cost per wafer }}{\text { Dies per wafer } \times \text { Yield }}$

$Dies per wafer \approx Wafer area/Die area$

$Yield =\frac{1}{(1+(\text { Defects per area } \times \text { Die area } / 2))^{2}}$

成品率

Defects per area：单位面积缺陷

Die area：模具面积

Defining Performance

Which airplane has the best performance? 从不同的方面进行考察。

[图片上传失败...(image-c230ba-1679189920085)]

Response Time and Throughput

Response time响应时间
- How long it takes to do a task(the time between the start and completion of a task)
Throughput吞吐量
- Total work done per unit time
- e.g., tasks/transactions/… per hour
How are response time and throughput affected by
- Replacing the processor with a faster version? 改善处理器
- Adding more processors to do separate tasks? 添加更多的处理器
- Queue ？采用排队机制，改善吞吐量
We’ll focus on response time for now…

Relative Performance

Define Performance = 1/Execution Time

“X is n time faster than Y”
$Performance _{X} / Performance _{Y} = Execution time _{Y} / Execution time _{X}=n$

Example: time taken to run a program

10s on A, 15s on B
Execution TimeB / Execution TimeA = 15s / 10s = 1.5
So A is 1.5 times faster than B

Measuring Execution Time

Elapsed time 消逝时间
- Total response time, including all aspects
  - Processing, I/O, OS overhead, idle time
- Determines system performance
CPU time（共享时,独自占用CPU时间）
- Time spent processing a given job
  - Discounts I/O time, other jobs’ shares
- Comprises user CPU time and system CPU
  time
- Different programs are affected differently by
  CPU and system performance

CPU Clocking

Operation of digital hardware governed(掌控) by a constant-rate clock （数字同步电路）

[图片上传失败...(image-31fbb2-1679189920085)]

Clock period: duration of a clock cycle
- e.g., 250ps = 0.25ns = 250×10^–12s
Clock frequency (rate): cycles per second
- e.g., 4.0GHz = 4000MHz = 4.0×10^9Hz

CPU Time

CPU Time = CPU Clock Cycles x Clock Cycle Time = $\frac{\text { CPU Clock Cycles }}{\text { Clock Rate }}$

Performance improved by

Reducing number of clock cycles
Increasing clock rate
Hardware designer must often trade off(折中)clock rate against cycle count

CPU Time Example

Computer A: 2GHz clock, 10s CPU time
Designing Computer B
- Aim for 6s CPU time
- Can do faster clock, but causes 1.2 × clock cycles
How fast must Computer B clock be?

$Clock Cycles _{A}= CPU Time _{A} \times Clock Rate _{A}$

= $10 \mathrm{~s} \times 2 \mathrm{GHz}=20 \times 10^{9}$

= $\frac{1.2 \times 20 \times 10^{9}}{6 \mathrm{~s}}=\frac{24 \times 10^{9}}{6 \mathrm{~s}}=4 \mathrm{GHz}$

Instruction Count and CPI

$Clock Cycles = Instruction Count \times Cycles per Instruction$

$CPUTime = Instruction Count \times CPI \times Clock Cycle Time$

$=\frac{\text { Instruction Count } \times \mathrm{CPI}}{\text { Clock Rate }}$

Instruction Count for a program
- Determined by program, ISA and compiler
Average cycles per instruction
- Determined by CPU hardware
- If different instructions have different CPI 指令具有不同CPI
  - Average CPI affected by instruction mix

CPI Example

Computer A: Cycle Time = 250ps, CPI = 2.0
Computer B: Cycle Time = 500ps, CPI = 1.2
Same ISA
Which is faster, and by how much?

[图片上传失败...(image-96f5ac-1679189920085)]

CPI in More Detail

If different instruction classes take different numbers 每指令类CPI不同，且指令出现频率不同

$\text { Clock Cycles }=\sum_{\mathrm{i}=1}^{n}\left(\mathrm{CPI}_{\mathrm{i}} \times \operatorname{Instruction~Count~}_{\mathrm{i}}\right)$

Weighted average CPI(平均CPI)

$\mathrm{CPI}=\frac{\text { Clock Cycles }}{\text { Instruction Count }}=\sum_{\mathrm{i}=1}^{\mathrm{n}}\left(\mathrm{CPI}_{\mathrm{i}} \times \frac{\text { Instruction Count }_{\mathrm{i}}}{\text { Instruction Count }}\right)$

CPI Example

Alternative compiled code sequences using instructions in classes A, B, C (三类指令)

[图片上传失败...(image-149eeb-1679189920085)]

Which code sequence executes the most instructions? sequence2

Which will be faster?

What is the CPI for each sequence?

[图片上传失败...(image-2f1cc7-1679189920085)]

Performance Summary

$\text { CPU Time }=\frac{\text { Instructions }}{\text { Program }} \times \frac{\text { Clock cycles }}{\text { Instruction }} \times \frac{\text { Seconds }}{\text { Clock cycle }}$

Performance depends on

Algorithm: affects IC(指令数), possibly CPI
Programming language: affects IC, CPI
Compiler: affects IC, CPI
Instruction set architecture: affects IC, CPI, Tc

Power Trends

[图片上传失败...(image-a6fd73-1679189920085)]

In CMOS IC technology

$\text { Power }=\frac{1}{2} \text { Capacitive load } \times \text { Voltage }^{2} \times \text { Frequency }$

Capacitive load:负载电容。

Reducing Power

Suppose a new CPU has

85% of capacitive load of old CPU
15% voltage and 15% frequency reduction

$\frac{P_{\text {new }}}{P_{\text {old }}}=\frac{C_{\text {old }} \times 0.85 \times\left(V_{\text {old }} \times 0.85\right)^{2} \times F_{\text {old }} \times 0.85}{C_{\text {old }} \times V_{\text {old }}^{2} \times F_{\text {old }}}=0.85^{4}=0.52$

The power wall (功率墙)
- We can’t reduce voltage further 可能低压泄露
- We can’t remove more heat 可能sleep
How else can we improve performance?

[图片上传失败...(image-4959f3-1679189920085)]

Constrained by power, instruction-level parallelism, memory latency（受到功率、指令级并行性、内存延迟的制约）

Multiprocessors（多核）

Multicore microprocessors
- More than one processor per chip
Requires explicitly parallel programming
- Compare with instruction level parallelism(e.g.流水线）
  - Hardware executes multiple instructions at once
  - Hidden from the programmer (程序员不可见)
Hard to do
- Programming for performance 编程难度增加
- Load balancing 负载均衡
- Optimizing communication and synchronization

A R M提供更多计算核心

[图片上传失败...(image-e4eed6-1679189920085)]

多核架构单位芯片面积提供更强算力，更符合分布式业务的需求

[图片上传失败...(image-f02cd6-1679189920085)]

A R M多核高并发优势，匹配互联网分布式架构

[图片上传失败...(image-433c68-1679189920085)]

随着多核A R M CPU的性能不断增强，应用领域不断扩展

[图片上传失败...(image-b9be8e-1679189920085)]

A R M服务器级别处理器一览

[图片上传失败...(image-a260b4-1679189920085)]

SPEC CPU Benchmark

Programs used to measure performance
- Supposedly typical of actual workload
Standard Performance Evaluation Corp (SPEC)
- Develops benchmarks for CPU, I/O, Web, …
SPEC CPU2006
- Elapsed time to execute a selection of programs
  - Negligible I/O, so focuses on CPU performance
- Normalize relative to reference machine（参考机器）
- Summarize as geometric mean of performance ratios
  - CINT2006 (integer) and CFP2006 (floating-point)

$\sqrt[n]{\prod_{\mathrm{i}=1}^{n} \text { Execution time ratio }_{i}}$

CINT2006 for Intel Core i7 920

[图片上传失败...(image-fc9b3f-1679189920085)]

SPEC Power Benchmark

Power consumption of server at different workload levels

Performance: ssj_ops/sec
Power: Watts (Joules/sec)

[图片上传失败...(image-ef762b-1679189920085)]

SPECpower_ssj2008 for Xeon X5650

[图片上传失败...(image-f46347-1679189920085)]

Pitfall(陷阱): Amdahl’s Law

Improving an aspect of a computer and expecting a proportional improvement in overall performance

$T_{\text {improved }}=\frac{T_{\text {affected }}}{\text { improvement factor }}+T_{\text {unaffected }}$

Example: multiply accounts for 80s/100s

Speedup(E)=1/{(1-P)+P/S}

Amdahl's law主要的用`途是指出了在计算机体系结构设计过程中，某个部件的优化对整个结构的优化帮助是有上限的，这个极限就是当S->时, speedup(E)= 1/(1-P);也从另外一个方面说明了在体系结构的优化设计过程中，应该挑选对整体有重大影响的部件来进行优化，以得到更好的结果。

Fallacy谬误: Low Power at Idle

Look back at i7 power benchmark

At 100% load: 258W
At 50% load: 170W (66%)
At 10% load: 121W (47%)

Google data center

Mostly operates at 10% – 50% load
At 100% load less than 1% of the time

Consider designing processors to make power proportional to load

Pitfall: MIPS as a Performance Metric

MIPS: Millions of Instructions Per Second

Doesn’t account for 考虑
- Differences in ISAs between computers
- Differences in complexity between instructions
$\begin{aligned} \text { MIPS } &=\frac{\text { Instruction count }}{\text { Execution time } \times 10^{6}} \\ &=\frac{\text { Instruction count }}{\frac{\text { Instruction count } \times \mathrm{CPI}}{\text { Clock rate }} \times 10^{6}}=\frac{\text { Clock rate }}{\mathrm{CPI} \times 10^{6}} \end{aligned}$
CPI varies between programs on a given CPU

Concluding Remarks

Cost/performance is improving

Due to underlying technology development

Hierarchical layers of abstraction

In both hardware and software

Instruction set architecture

The hardware/software interface

Execution time: the best performance measure

Power is a limiting factor

Use parallelism to improve performance