See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/342867217

Comparative Study on CPU, GPU and TPU

Article · May 2020

DOI: 10.21742/IJCSITE.2020.5.1.04

CITATIONS READS
5 2,380

2 authors, including:

Ch Sekhar
Vignan’s Institute of Information Technology
46 PUBLICATIONS 86 CITATIONS

SEE PROFILE

All content following this page was uploaded by Ch Sekhar on 07 September 2021.

The user has requested enhancement of the downloaded file.

 International Journal of Computer Science and Information Technology for Education
Vol.5, No.1 (2020), pp.31-38
http://dx.doi.org/10.21742/ijcsite.2020.5.1.04

Comparative Study on CPU, GPU and TPU

P Siva Raj1 and Ch. Sekhar2

Dept. of Comp. Science & Engineering, Vignan’s IIT (Autonomous), AP, India
1
[email protected], [email protected]

Abstract
The recent trend in the hardware technology changed as per need of the processing of
huge amount of data receiving in the form of text, image, videos and graphical data. In this
paper, we are going to discuss the evolutionary and need for change in hardware such as
central processing units, graphical processing units and tensor processing units. Briefly
explain the concept of each with architecture. Providing a comparative analysis of each unit.

Keywords: CPU, GPUs, TPUs, Processing units, Tensor

1. Introduction
The present trend of usage of the computer system we changed allots to last two decades. Past
10 years tremendous change came to the user perspective. The user used the system not only
for the processing of standard data, used for all kind of data accesses of the different forms
such as text, image, video, gaming, structured, semi-structured and unstructured data. The
usage tendency changed allot of the computer. Initially, we focused more on the processing
of flat data by using a CPU based system with single or multi-core processors. Later based the
need and processing of graphical data, we started working with GPUs.
Graphical processing unit came into the market in the year 1999 by the manufacturer
NVIDIA. Graphical Processing unit simply shortened as GPUs. These have simply settled
capacity gadgets, implying with the intention of explicitly progression phases of illustrations
pipeline. For example, zenith and pixel shavers, however, they have developed into
progressively adaptable coding processors. up to date, Graphical Process Units are completely
programmable numerous core chips worked around a variety of comparable processors [1].

Figure 1. a) Central processing unit b) Graphical processing unit c) Tensor unit

Article history:
Received (February 11, 2020), Review Result (March 15, 2020), Accepted (April 16, 2020)

Print ISSN: 2205-8370, eISSN: 2207-5372 IJCSITE

Copyright ⓒ 2020 Global Vision Press (GV Press)
Comparative Study on CPU, GPU and TPU

2. Centralized Processing Unit (CPU)

CPU is the shorten name of Central Processing Unit, is the hardware with a combination of
electronics, which functions as brains of the PC that play out the essential number-crunching,
sensible, control and information/yield activities indicated by the instructions of a PC
program [2]. As far as figuring power, the CPU is an essential component of a PC system. A
massive collection of gadgets uses the CPU, with work locale, tablets, PC and workstations,
Mobile phones even your smart TV.
The 3 essential elements or components of CPU, the operations of logical operations,
arithmetic operations and the centralized unit, using the internal registers up to some size of
data based on the bit size of the processor. Control Unit (CU), which removes directions from
memory and disentangles and executes them, approaching the ALU when fundamental. Also,
Memory administration unit, which is accessible in most top of the line chip to interpret
legitimate locations into physical RAM addresses.
Presently, the CPU accompanies single and multi-core variation. Multi-core means having
more than one processor units functioning one subsequently to the other implies that the CPU
can oversee over one guideline consistently, all the while, radically enhancing execution. [6]
The most popular manufacturers are AMD and INTEL work areas, PCs, and servers. At the
same time, the enormous models of mobile phone used processor manufacturers are
QUALCOMM, APPLE and NVIDIA.

Figure 2. Multi-core processor

The processor can produce utilizing diverse advances - Single-centre CPU and more core
processor. As per processors can be isolated into three kinds multiprocessors, multithreaded
processors and multi-core processors [3]. There are new patterns in the CPU producing
industry which depend on the possibility that while at the same time clock velocities must be
expanded as far as possible and there is breaking point to number of electronic parts to be
utilized as a part of a center. Numerous different advancements are there to speed things up
and open routes for better and all the greater focal handling units [3]. When we can’t build the
execution of CPU moreover by changing its running recurrence, at that point, an innovation
called multi-core engineering makes a difference. In multi-core design, we can put over one
centre on a solitary silicon kick the bucket. This new way to deal with improve the speed

32 P Siva Raj and Ch. Sekhar

 International Journal of Computer Science and Information Technology for Education
Vol.5, No.1 (2020), pp.31-38

accompanied some extra advantages like better execution, better power administration and
better cooling as the multicenter processors keep running at a lower speed to disseminate less
warmth. It additionally has a few inconveniences like existing projects should be revamped
according to new design. On the off chance that we don’t compose programs with an
exceptional centre for circling parallel centres, we won’t get the favourable position of multi-
cores.

3. Graphical Processing Unit (GPU)

The need for GPUs advanced to contain extensive quantities of similar threads and
numerous cores? The main thrust keeps on being the execution of the ongoing illustrations
required to render intricate, high-determination 3D scenes at reasonable edge rates for
amusements [4].
Rendering top quality designs scenes is an issue with gigantic characteristic parallelism.
An illustrations software engineer composes with one thread program that can provide one
pixel, and the GPU runs various cases of this thread in similar—drawing various pixels in
parallel. Illustrations programs, written in shading languages, for example, Cg or High- Level
Shading Language (HLSL), in this manner scale thread forwardly finished a wide variety of
string.

Figure 3. Evolution of the graphics programming on the GPUs: Courtesy [1]

As we come into the period of GPU figuring, requesting applications with considerable
parallelism progressively utilize the significantly corresponding figuring abilities of GPUs to
accomplish unrivaled execution furthermore, proficiency. Today GPU processing empowers
applications that we beforehand attention infeasible because of the lengthy finishing epoch.
Faster and quick possible by the GPU’s from a configurable illustration workstation to a
programmable a parallel central processing unit. The pervasive GPU in each PC, workstation,
work area, what’s more, the workstation is a many-centre multithreaded multiprocessor that
exceeds expectations at the two illustrations and registering applications. The present GPUs
utilize several parallel processor centres executing several thousands of parallel strings to
fathom extensive issues having generous natural parallelism quickly. They’re currently the

Copyright © 2020 Global Vision Press (GV Press) 33

Comparative Study on CPU, GPU and TPU

most unavoidable hugely parallel preparing stage ever accessible and also the most financially
savvy.
During the period of 1990s, there were no GPUs, and Video Graphic Adapter controllers
produced 2D designs show for PCs to quicken graphical UIs [4]. During the year 1997’s,
NVIDIA came with the 3D as single-chip called RIVA 128, quickening agent for recreations
and 3Dperception applications, customized with Microsoft Direct3D and OpenGL.
Developing to present-day GPUs included including programmability incrementally—from
settled work pipelines to microcode processors, configurable processors, programmable
processors, what’s more, adaptable parallel processors.
The GPU comprises of a variety of Streaming Multiprocessors (SM), every one of which is
fit for supporting a great many co-inhabitant simultaneous equipment strings, up to 2048 on
current design GPUs. All string administration, including creation, booking, and hindrance
synchronization is performed completely in equipment by the SM with basically zero
overhead. To productively deal with its extensive string populace, the SM utilizes a SIMT
(Single Instruction, Multiple Thread) engineering.

(Image copied from NVIDIA_CUDA_Tuotorial_No_NDA_Aprs08.pdf)

Figure 4. Architecture of GPU

4. Tensor Processing Units (TPU)

A Tensor is an n-dimensional network. This is the fundamental unit of activity in with
TensorFlow, the open-source AI system propelled by Google Brain. A Tensor is practically
equivalent to a NumPy exhibit and in truth utilizes Numpy. As indicated by their
documentation, it is “NumPy is the key bundle for logical figuring with Python. It contains in
addition to other things a ground-breaking N-dimensional cluster object Clusters are the key
information structures utilized by AI calculations. Increasing and taking cuts from clusters
takes a great deal of CPU clock cycles and memory. So Numpy was composed to make
composing code to do that simpler. GPUs now make those tasks run quicker than the other
models.
Google began searching for a way to support neural networking for the development of
their services such as voice recognition Using existing hardware; they would require twice as
many datacenters.

34 P Siva Raj and Ch. Sekhar

 International Journal of Computer Science and Information Technology for Education
Vol.5, No.1 (2020), pp.31-38

Figure 5. Architecture of tensor

Google Google
Applicati Applicat
ons ions

TensorFlow

Stream Executor API

User Space

Kernel

TPU

Figure 6. Application stack of tensor

 Development of a new architecture instead Norman Jouppi begins work on a new

architecture to support TensorFlow
 FPGA’s were not power-efficient enough
 ASIC design was selected for power and performance benefits

Copyright © 2020 Global Vision Press (GV Press) 35

Comparative Study on CPU, GPU and TPU

 The device would execute CISC instructions on many networks

 The device was made to be programmable, but operate on matrices instead of
vector/scalar
 The resulting device was comparable to a GPU or Signal Processor
Based on the above study it the following are comparative analysis given as

Table 1. Comparative analysis

Parameter CPU GPU TPU

10’s operations per 10-103 operations per Up to 128000
Performance
Cycle Cycle operations
Dimension of data Unit of 1 x 1 Unit of 1 x N Unit of N x N
Graphical
Usage Normal Programming Machine Learning
Programming
Intel, AMD, IBM,
Manufacturers NVIDIA,AMD Google
Samsung
Cost of Machine 10-15 $ 150-200 $ 350-450$

Figure 7. Single processor vs multi-core processor performance

Figure 8. Performance analysis of GPUS vs CPU

36 P Siva Raj and Ch. Sekhar

 International Journal of Computer Science and Information Technology for Education
Vol.5, No.1 (2020), pp.31-38

5. Conclusion
In this paper, we elaborated the basics architecture views of various processing
technologies used in the computer. Such as CPU, GPU and TPU, when we want work parallel
task programming we can use Central Processing Units (CPUs), data-parallel over the
processors we will choose the Graphics Processing Units (GPUs) [5][7]. Besides their
traditional use as graphics coprocessors, the GPUs have been used in recent years for general
purpose computations, too. The rapid progress of graphics hardware led to widespread use in
both scientific and profitable applications. Numerous papers report high speedups in various
domains. This paper presents an effort to bring GPU computing closer to programmers and a
wider community of users. The most popular application programming interface, i.e. API, is
with the combination of CUDA and NVIDIA to provide general purpose application of GPUs.

References
[1] G. Sukhdev Singh, “Comparison of single-core and multi-core processor”, IJARCSSE, vol.6, no.6, pp.423-
424, (2016)
[2] Hemsoth N., First In-Depth Look at Google’s TPU Architecture, https://www.nextplatform.com/2017/04/05/
first-depth-look-googles-tpu-architecture, (2017)
[3] Geekboots, “CPU vs GPU vs TPU,” https://www.geekboots.com/story/cpu-vs-gpu-vs-tpu, (2020)
[4] N. John and D. William, “The GPU computing era”, Micro, IEEE Computer Society, vol.30, pp.56-69, (2010)
[5] M. Misic and D. Durdevic, “Evolution and trends in GPU computing”, pp.289-294, (2012)
[6] Tatourian A., “NVIDIA GPU Architecture & CUDA Programming Environment,” https://tatourian.blog
/2013/09/03/nvidia-gpu-archit>ecture-cuda-programming-environment/, (2013)
[7] R. Abinash, X. Jingye, and C. Masud., “Multi-core processors: A new way forward and challenges”, ICM,
pp.454-457, (2008)

Copyright © 2020 Global Vision Press (GV Press) 37

 Comparative Study on CPU, GPU and TPU

This page is empty by intention.

38 P Siva Raj and Ch. Sekhar

View publication stats