Advanced Investigation and Comparative Study of Graphics Processing Unit-queries Countered

GPU, Graphics Processing Unit, is the buzz word ruling the market these days. What is that and how has it gained that much importance is what to be answered in this research work. The study has been constructed with full attention paid towards answering the following question. What is a GPU? How is it different from a CPU? How good/bad it is computationally when comparing to CPU? Can GPU replace CPU, or it is a day dream? How significant is arrival of APU (Accelerated Processing Unit) in market? What tools are needed to make GPU work? What are the improvement/focus areas for GPU to stand in the market? All the above questions are discussed and answered well in this study with relevant explanations.


INTRODUCTION
What is GPU (Graphics Processing Unit)?Expanded as Graphics Processing Unit, the name itself is self-revealing what it is meant for.It is specially used for increasing the graphics handling ability of a computer.But, this brief is not sufficient.It is also referred to be as VPU (Visual Processing Unit) which is intended to swiftly direct and alter memory to speed up the creation of images to be displayed.Experts say this way, GPU (Fatahalian and Houston, 2008;Brodtkorb et al., 2010;Davidson and Owens, 2010) is a solo chip processor used to administer and enhance the performance of video and graphics.Usage of GPU is not only restricted to PC, but also be extended to mobile phones, display adaptors', work stations, Medical imaging, Gaming etc.It is inevitable to check the utility areas of GPU.GPUs has got a wide range of application areas which include, Enterprise applications, Medicinal imaging, Gaming, Augmented reality applications, Virtual reality applications and many more.These days' people who use PCs, Televisions, Laptops, Mobile phones want a rich experience of visualizing things on the screen.It is made possible through GPU.The next question to be answered is how GPU is different from the CPU?The following image could be familiar to the readers; the NFS games would be looking and feeling awesome with GPUs in place.Figure 1 is gist of the rich graphics supported by GPU.Main aim of drafting this study is to get a clear cut idea on GPUs.

MATERIALS AND METHODS
How is it different from a CPU? NVIDIA makes this interesting claim.CPU is always regarded as brain of PC, whereas we would say GPU enhances the CPU.So GPU can be definitely certified as Soul of PC.This statement uncovers the importance paid to GPU.
As shown above in the Fig. 2, CPU and GPU both are composed of millions of transistors.The architectural details of the GPU are maintained as trade secret, where we could understand the architecture to a decent extent with the articles available online.CPU devotes most of its transistors towards control, DRAM and cache while assigning a limited amount of transistors to computational tasks and hence limiting the computational ability.GPUs are different in this aspect.GPUs dedicate most of the transistors towards the computational part i.e., ALU which will ultimately increase the computational ability of the processor.And yes, this makes GPU the unanimous choice for applications with undemanding control flow and high numerical intensity.Technically speaking the above claimed statements can be little better re-phrased.CPU is actually built with few cores with lots of cache memory that would be capable of handling only a few software threads at a time.But, GPU on the other hand is built with hundreds of cores which could handle thousands of threads at an instance.The highly admirable view of GPU is about its ability to achieve high acceleration with being more power and cost efficient compared to CPU.CPU can be seen as small group of very smart people who can swiftly do any task  given to them at a specified short duration of time.But, GPU could be a large group of not smart people, who are competent individually, but when made to work in a team, they outperform CPU.
History of GPUs: GPU has been evolving consistently over a last and half decades (Harris and Göddeke, 2002).The first and foremost 3D graphics all emerged with early display regulators referred to be as Video Shifters or VAG (Video Address Generators).They were designed to act as a pass-through amongst the core processor and the display unit.The output would be serial bit mapped and synchronized properly based on color, luminance, horizontal and vertical composition etc.
Motorola, the genius in the company revealed the MC6845 VAG (Wikipedia, 2010a).This was one of the most frequently used adapters by IBM, Apple and other giants who wanted the display to be having higher clarity.Since, the business was grooming up well in the VAG sector, Motorola revealed one more addition to the VAG sector, MC6847.It was used in the first generation PCs which did not exclude TRS-80.Figure 3 shows the IBM monochrome display adapter.
Intel was observing Motorola's move into the market and they have understood it is time for Intel to jump in to the market of Graphics adaptors.In 1978, iSBX275 VGCMB (Video Graphics Controller Multimodule Board) was unveiled.Drawing of character bitmaps, arcs, lines along with rectangles are all well augmented with the help of Direct Memory Access (DMA).
In 1980s, VGA (Video Graphics Array) is the one which was used before the advent of GPUs to handle graphics in the PCs.Since VLSI techniques and ability to handle transistors were both developing, VGA's quality and coverage has been enhanced to a greater extent in handling the images and video.
During the year 1985, there came the change which changed the way things looked in the past.Three researchers from Hong Kong Kwok Yuan Ho, Lee Lau and Benny Lau, formed Array Technology Inc. which is referred as ATI Technologies later.The product released by ATI immediately after its inception is OEM Color Emulation Card.It was useful to get the monochrome green, amber or white text in contradiction of black background to a Monitor.This was the path breaking invention in this field.Following 1985Following , 1986 was is a very notable year where Texas Instruments unveiled TMS34010 which is a processor with having on chip graphical abilities.Though it had a good graphics processing ability and an instruction set supporting it, it also was used to handle general purpose tasks.TIGA-Texas Instruments Graphics Architecture, which later ruled the graphics world has TMS 34010 as base.In short this decade of 1980s had been wonderful for the graphics world.
Next came the most notable years of invention in the field of Graphics cards.2D GUIs acceleration has started booming throughout this decade.APIs started finding its place in the market.3D graphics with support from CPU has really started working out well and it was most common within gamers.Many organizations started working towards the low cost 3D graphics card.As ever, the first few of them failed to attain the glory.Few to mention are S3's ViRGE, ATI's Rage etc.Though they failed, they obviously mean to be the forefathers of the current day GPUs.As the days progressed and brains worked faster, things started happening.Video, 2D GUI, 3D features etc., are all integrated in under one roof of silicon.
In 1990's, came the name which has made revolution in the market of Graphics, OpenGL.It was not a great performer in the early stages like any other invention, with performance flaws.Then the issues were fixed and it started contributing well.There arisen among topographies offered in hardware and those The GPUs which have come to existence in the market today is based on Graphics pipeline and the same has been presented above in diagrammatic format.It actually models the stages that the graphical data really passes through.The stages is actually composed of a combination of blend of hardware and software where hardware is GPU Cores and the software can be something we have already referred to in this study as OpenGL or DirectX.There can be a question now asked that is this pipeline commonly followed by all the GPU manufacturers?Yes, it is precisely followed by all the manufacturers of GPU as NVIDIA ( 2010) and Owens et al. (2008) and ATI.The transformation is simply admirable as the coordinates from the 3D space gets transformed to 2D space in the screen.The generation wise look up is given in Fig. 5.
Can GPU be a CPU?First, GPGPU (Li et al., 2009) is the term coming to picture now.To answer this question, GPGPU is general purpose GPU.Indeed, one can simply see it as a GPU which will/can act as a CPU.
As we all know, the fabrication methods and technologies are advancing every now and then.This will make the major changes in the performance and cost.CPUs generally are constructed with having only performance in mind.Many or most of the transistors are moved towards providing non mathematical or computational tasks as caching.But, the GPU (Buck, 2010) is different in this aspect.It is parallel in nature and that permits the GPUs to have more transistors used for computation which will help in achieving high mathematically dense operations, without having need to increase the number of transistors.To summarize quickly, GPUs are: • Powerful: NVIDIA GeForce 6800 (NVIDIA, 2011(NVIDIA, , 2012) ) Ultra is capable of reaching 35.2 GB/sec of memory bandwidth.• Less expensive: Above said processor is just $417, so definitely it is cheaper and cost effective.• Flexible: Yes, it is really flexible without any much of operating system and other similar dependencies.
• Programmable: This is the highlight.It is expandable.It makes the life and creativity live.
How good/bad it is computationally when comparing to CPU? CPU is seen to be computationally less powered comparing GPU.It is because of the following reasons: • CPU is restricted with having one thread on CPU core whereas in GPU there could be as many as several hundred threads on one GPU core.• Most importantly, in CPU one thread has to take a pause for the next thread to run.In case of GPU it is different.It can queue at least thousands of threads and has to ability to hang some threads to run others.• GPU has a policy which lets all threads to do the same thing which would fetch an efficient execution.• Data parallelism remains the key factor for the success of GPU's which provide very high performance.
• High performance is provided with having above conveyed is unique and it has a programming model which stays different from the legacy CPU (Micikevicius, 2010a) programming model.• GPUs are able to deliver because of the number of transistors present in it.There is a difference when comparing it with CPU.GPUs use smaller transistors when compared to the CPU.And this makes an obvious difference in the computation and processing ability.• GPU is normally designed and expected to compute operations of the order of millions which is not in the case of CPU.Also GPUs (Wikipedia, 2010b) are capable of handling large textures as feed.Another point to observe which could seem to be quiet contradictory while reading is the latency related.Latency is for GPU is very high when having the comparison done with the CPU.Reason is being very simple, latency which normally would not exceed milli seconds are certainly not traceable by the human eyes.They will appear as such there is no latency at all.• The memory structure is not the same and it is overt that they can't really be.A simple diagrammatic explanation would reveal this fact.
The cache memory distribution is the area there is a significant difference GPU has given similar emphasize on the L1 and L2 wherein for CPU it is not the case.The cache hit rates are appearing to be comparatively much lesser for GPUs than the case of CPUs. Figure 6 shows the memory of CPU and GPU.Unified Address Space for CPU and GPU-Pointers can now be freely passed between CPU and GPU, hence enabling zero-copy.Fully coherent memory between CPU and GPU-Cache coherency is maintained.GPU uses pageable system memory via CPU pointers-GPU can take advantage of the shared virtual memory between CPU and GPU and pageable system memory can now be referenced directly by the GPU, instead of being copied or pinned before accessing.System integration 2015 Carrizo APU GPU compute context switch-allowing a multi-tasking environment and also faster interpretation between applications, compute and graphics.GPU graphics pre-emption-Long-running graphics tasks can be pre-empted so processes have low latency access to the GPU.Quality of Service-In addition to context switch and pre-emption, hardware resources can be either Equalized or prioritized among multiple users and applications.source software.It is equally good computationally when compared to other options available in the market.Support is rendered to CPUs, GPUs and DSPs which is collar lifting fact on the OpenCL.A comparison is always nice and the below table provides the same between the OpenCL and CUDA with respect to API (Table 1 and 2).
The AMD Accelerated Processing Unit, formerly known as Fusion, is a series of 64-bit microprocessors from AMD designed to act as a CPU and graphics accelerator (GPU) on a single chip (Table 3).

RESULTS AND DISCUSSION
GPUs presently are not interacting much with cloud kind of infrastructure.But, in future it could be the need of the hour to make GPU (Micikevicius, 2010b) work in coherence with cloud.It is definitely possible to have a better bottleneck reduced application support in the cloud with GPU.Technology grows towards the innovation and following are simple examples which we are eyeing in front of us in the recent past: • CASS is one of the leading players coming up with Embedded GPU platforms, Hoopoe GPU cloud computing and CLIPP.• NVIDIA is focusing towards the deep learning algorithms which definitely would benefit in producing better artificial intelligent systems.• AMD has started to work on next generation processors named Carrizo which is expected to be a boom in the market of graphics.

CONCLUSION
We the authors have made efforts in understanding of the trends in GPU and the same has been presented in the study.The paper has clarified on what a GPU is and how is it different from a CPU.Also is it a solid replacement for legacy CPU or it is just hype in the modern era?How is the market reacting towards this innovation and what are the companies working to make the revolution happen?All these have been discussed and clarity has been given to the same.Also where can the GPU era head towards is projected in the paper.The same paper can be further crafted with giving more details on all the individual GPUs available in the market.

•
Comparing instruction set is the next task.Earlier, the life was tedious where each instruction is limited to perform one operation.Means, every operation needs one instruction for the destiny to be reached.It was referred to be as CISC and later, the trend changed where one instruction can be used at different occasions to perform/achieve different tasks.It is termed as RISC.Since the navigation happens towards more of general purpose computing with the GPU the following supports are also made available and it is included: o Increased processing ability with included floating point calculations.o Uninformed reading option from memory is also provided.o Options to use Loops and Conditional constructs made the GPUs more GP.• These above points are all taking the drift towards GPU becoming more like CPU with inclusion of CPU like elements as loops, conditional constructs, floating point capabilities etc. SDK tool kits are available for variety of platforms starting from most used Linux, Windows and more sophisticated MAC OS for Apple.Well, is there a competitor from any other source?Yes, there is a competitor.It is named as OpenCL and it is open