GPU vs. CPU for dummies (like me)
When I started the fast.ai course and was just learning what the heck the notebooks are (places where you can code, something like Colab and Gradient).
Our mentors were saying something like this: “Don’t use CPU, switch to GPU. Only if you use GPU, your model can work fast; the CPU approach can take hours when GPU can make it within some minutes.”
To me, it sounded like: “Please don’t touch this red bottom, because the plant will be blown away; touch the blue one instead, and all will be ok.”
If you are just starting and the phrases like the one above sound strange to you, let me try to explain in a way that I explained it to myself (dear experienced developers, please feel free to comment if you see that I’m making mistakes in this post).
The best explanation of CPU that I found:
The CPU (central processing unit) of a computer is often called the “brain” of a computer. It’s a collection of millions of transistors that can be manipulated to perform an awesome variety of calculations. A standard CPU has between one and four processing cores clocked anywhere from 1 to 4 GHz.
A CPU is powerful because it can do everything. If a computer is capable of accomplishing a task, it’s because the CPU can do it. Programmers achieve this through broad instruction sets and long feature lists shared by all CPUs.
So, let’s call your CPU a hero. Some kind of the smartest or the biggest badass in your office or team. Somebody who can manage the toughest tasks.
I decided to choose Elon Musk for this role (what a cliche, right?).
So, Musk is CPU.
What about GPU? Here is the explanation?
A GPU (graphics processing unit) is a specialized type of microprocessor. It’s optimized to display graphics and do very specific computational tasks. It runs at a lower clock speed than a CPU but has many times the number of processing cores.
You can almost think of a GPU as a specialized CPU that’s been built for a very specific purpose. Video rendering is all about doing simple mathematical operations over and over again, and that’s what a GPU is best at. A GPU will have thousands of processing cores running simultaneously. Each core, though slower than a CPU core, is tuned to be especially efficient at the basic mathematical operations required for video rendering. This massive parallelism is what makes GPUs capable of rendering the complex 3D graphics required by modern games.
In other words, GPU is an army of goofy creatures doing the same tasks, but there are a lot of them:
A quote from another website once again:
A GPU can only do a fraction of the many operations a CPU does, but it does so with incredible speed. A GPU will use hundreds of cores to make time-sensitive calculations for thousands of pixels at a time, making it possible to display complex 3D graphics. However, as fast as a GPU can go, it can only really perform “dumb” operations.
On Fair Comparison between CPU and GPU
As a noob newbie Computer Science researcher, it is always fun and rewarding to watch people discussing about our research papers somewhere on the Internet. Besides some obvious implications (e.g., fresh perspectives from practitioners on the research subjects), it is a strong indicator that my paper was not completely irrelevant junk that get published but nobody really cares about it. Today, I came across a tweet thread about our SSLShader work, or more specifically, about the RSA implementation on GPUs, which is what I was responsible for, as a second author of the paper. Soon I found a somewhat depressing tweet about the paper.
”[…] the benchmark methodology is flawed. Single threaded CPU comparison only.”
Ugh… “flawed”. Really? I know I should not take this personal, but my heart is broken. If I read this tweet at night I would have completely lost my sleep. The most problematic(?) parts of the paper are as follows, but please read the full paper if interested:
The figure, table, and text compare the RSA performance between the GPU (ours) and CPU (from Intel IPP) implementations. The CPU numbers are from a single CPU core. Do not throw stones yet.
To be fair, I think the negative reaction is pretty much reasonable
(after having a 2-hour meditation)
. There are a bunch of academic papers that make clearly unfair comparisons between the CPU and GPU implementations, in order for their GPU implementations to look shinier than actually are. If my published paper was not clear enough to avoid those kinds of misunderstanding, it is primarily my fault. But let me defend our paper a little bit, by explaining some contexts on how to make a fair comparison in general and how it applied to our work.
How to Make Fair Comparisons
It is pretty much easy to find papers claiming that “our GPU implementation shows an orders of magnitude speedup over CPU”. But they often make a comparison between a highly optimized GPU implementation and an unoptimized, single-core CPU implementation. Perhaps one can simply see our paper as one of them. But trust me. It is not what it seems like.
Actually I am a huge fan of the ISCA 2010 paper, “Debunking the 100X GPU vs. CPU Myth”, and it was indeed a kind of guideline for our work to not repeat common mistakes. Some quick takeaways from the paper are:
100-1000x speedups are illusions. The authors found that the gap between a single GPU and a single multi-core CPU narrows down to 2.5x on average, after applying extensive optimization for both CPU and GPU implementations.
The expected speedup is highly variable depending on workloads.
For optimal performance, an implementation must fully exploit opportunities provided by the underlying hardware. Many research papers tend to do this for their GPU implementations, but not much for the CPU implementations.
In summary, for a fair comparison between GPU and CPU performance for a specific application, you must ensure to optimize your CPU implementation to the reasonably acceptable level. You should parallelize your algorithm to run across multiple CPU cores. The memory access should be cache-friendly as much as possible. Your code should not confuse the branch predictor. SIMD operations, such as SSE, are crucial to exploit the instruction-level parallelism.
(In my personal opinion, CPU code optimization seems to take significantly more efforts than GPU code optimization at least for embarrassingly parallel algorithms, but anyways, not very relevant for this article.)
Of course there are some obvious, critical mistakes made by many papers, not addressed in detail in the above paper. Let me call these three deadly sins.
Sometimes not all parts of algorithms are completely offloadable to the GPU, leaving some non-parallelizable tasks for the CPU. Some papers only report the GPU kernel time, even if the CPU runtime cannot be completely hidden with overlapping, due to dependency.
More often, many papers assume that the input data is already on the GPU memory, and do not copy the output data back to the host memory. In reality, data transfer between host and GPU memory takes significant time, often more than the kernel run time itself depending on the computational intensity of the algorithm.
Often it is assumed that you always have large data for enough parallelism for full utilization of GPU. In some online applications, such as network applications as in our paper, it is not always true.
While it is not directly related to GPU, the paper “Twelve ways to fool the masses when giving performance results on parallel computers” provides another interesting food for thought, in the general context of parallel computing.
What We Did for a Fair Comparison
Was our CPU counterpart was optimized enough?
We tried a dozen of publicly available RSA implementations to find the fastest one, including our own implementation. Intel IPP (Integrated Performance Primitives) beat everything else, by a huge margin. It is heavily optimized with SSE intrinsics by Intel experts, and our platform was, not surprisingly, Intel. For instance, it ran up to three times faster than the OpenSSL implementations, depending on their versions (no worries, the latest OpenSSL versions runs much faster than it used to be).
Why show the single-core performance?
The reason is threefold.
RSA is simply not parallelizable, at a coarse-grained scale for CPUs. Simply put, one RSA operation with a 1k-bit key requires roughly 768 modular multiplications of large integers, and each multiplication is dependent on the result of the previous multiplication. The only thing we can do is to parallelize each multiplication (and this is what we do in our work). To my best knowledge, this is true not only for RSA, but also for any public-key algorithms based on modular exponentiation. It would be a great research project, if one can derive a fully parallelizable public-key algorithm that still provides comparable crypto strength to RSA. Seriously.
The only coarse-grained parallelism found in RSA is from Chinese Remainder Theorem, which breaks an RSA operation into two independent modular exponentiations, thus runnable on two CPU cores. While this can reduce the latency of each RSA operation, note that it does not help the total throughput, since the total amount of work remains the same. Actually IPP supports for this mode, but it shows lower throughput than the single-core case, due to the communication cost between cores. Fine-grained parallelization of each modular multiplication on multiple CPU cores is simply a disaster. Even too obvious to state.
For those reasons, it is best for the CPU evaluation to run the sequential algorithm on individual RSA messages, on each core. We compare the sequential, non-parallelizable CPU implementation performance with the parallelized GPU implementation performance. This is why we show the single-core performance. One can make a rough estimation for her own environment from our single-core throughput, by considering the number of cores she has and the clock frequency.
In our paper, we clearly emphasized several times that the performance result is from a single core, not to be misunderstood as a whole CPU or a whole system (our server had two hexa-core Xeon processors). We also state that how many CPU cores are needed to match the GPU performance. And finally, perhaps most importantly, we make explicit chip-level comparisons, between a GPU, CPUs (as a whole), and a security processor in the Discussion section.
What about the three deadly sins above?
We accounted all the CPU and GPU run time for the GPU results. They also include memory transfer time between CPU and GPU and the kernel launch overheads.
Our paper does not simply say that RSA always runs faster on GPUs than CPUs. Instead, it clearly explains when is better to offload RSA operations to GPUs and when is not, and how to make a good decision dynamically, in terms of throughput and latency. The main system, SSLShader, opportunistically switch between CPU and GPU crypto operations as explained in the paper.
In short, we did our best to make fair comparisons.
Time for Self-Criticism: My Faults in the Paper
Of course, I found that myself the paper was not completely free from some of common mistakes. Admittedly, this is a painful, but constructive aspect of what I can learn from seemingly harsh comments on my research. Here comes the list:
Perhaps the most critical mistake must be “In our evaluation, the GPU implementation of RSA shows a factor of 22.6 to 31.7 improvement over the fastest CPU implementation”. IN THE ABSTRACT. Yikes. It should have clearly stated that the CPU result was from the single-core case, as done in the main text.
Our paper lacks the context described above: “why we showed the single-core CPU performance”.
The paper does not explicitly say about what would be the expected throughput if we run the non-parallelizable algorithm on multiple CPU cores, individually. Clearly (single-core performance) * (# of cores) is the upper bound, since you cannot expect super-linear speedup for running on independent data. However, the speedup may be significantly lower than the number of cores, as commonly seen in multi-core applications. The answer was, it shows almost perfect linear scalability, since RSA operations have so small cache footprint that each core does not interfere with others. While the Table 4 implied it, the paper should have been explicit about this.
The graphs above. They should have had lines for multi-core cases, as we had done for another research project. One small excuse: please blame conferences with page limits. In many Computer Science research areas, including mine, conferences are primary venues for publication. Not journals with no page limits.
Gpu Vs Cpu What S The Difference A Guide
The GPU or graphics processing unit is a microchip component designed to create and render images and videos on computers and video game consoles. GPUs can be found on most computer systems with screens, including computers, mobile phones, tablets, and gaming consoles. GPUs are the most crucial hardware components when it comes to reproducing graphics. They take instructions and 3D graphics data from the CPU and process them into the screen. This is why they are also called graphics cards or video cards. GPUs would grow into this pre-determined role and evolve from displaying only a few characters to displaying photorealistic images with excellent frame rates. Furthermore, GPUs are composed of smaller units called arithmetic logic units (ALU) units, the grouping of which creates a variety of weaker cores. Generally speaking, GPUs require fewer ALUs compared to CPUs. The presence of a multitude of ALU units allows the GPU to process large volumes of data simultaneously. There are two types of GPUs available: dedicated graphics cards and integrated graphics processing units. Dedicated GPUs are often sold and bought separately and placed on a special slot next to the central processing unit (CPU). On the other hand, integrated GPUs are often pre-assembled with the motherboard. The ability to process multiple computations simultaneously makes GPUs the ideal candidate for processing large volumes of data, and its abilities extend far beyond graphics processing. In fact, the technology is being adopted into a wide selection of industries, including machine learning, robotics, statistics, linear algebra, medical research, and engineering, to name a few. Also, read about artificial intelligence and machine learning to find out about their similarities and nuances.
What Is CPU?#
The CPU or the central processing unit is the command center or main coordinator of a computer system. The CPU receives data from software, executes commands, and relays commands to other computer software components. Occasionally, the CPU would pass on mathematically complicated data (e.g., graphics) to the GPU for processing. CPUs are composed of smaller units called cores. Each core contains a combination of ALUs, a control unit, and registers. Each core can handle a single task simultaneously, usually on a sequential basis as the commands are executed. Older CPUs only had a single core that could process one task at a time. Nowadays, though, CPUs have anywhere between two to eighteen cores. These numbers are already quite a lot for CPUs, but they pale in comparison to the number of cores on GPUs. But then again, CPU cores are still far more powerful than GPU cores. Some CPUs have hyper-threading abilities. Hyper-threading is when a single physical CPU core appears as two virtual cores to the operating system. These virtual cores can then process separate commands simultaneously, and it also allows them to borrow resources from other cores. But while CPUs are commonly associated with computers and laptops, their use is not limited to computers alone. Most devices that run programs, such as phones, smartwatches, consoles, tablets, e-readers, and televisions, also use them. Also, read this comparative analysis of top CPU brands Intel vs. Ryzen to help you find the best picks for your PC.
GPU vs CPU: What’s the Difference?#
GPUs and CPUs are both silicone-based microprocessors. These microprocessors complement each other when it comes to allowing a computer to run. However, GPUs and CPUs accomplish different tasks and have different processes for accomplishing these tasks. These differences fall right into the heart of the GPU vs CPU discussion and shed light on the inner workings of both components.
Architecture#
The hardware for GPUs and CPUs look alike on the outside, but their inner architecture is different. CPUs contain a few complicated cores that number between two to 18. These cores are designed to break down tasks into smaller pieces, allowing multiple cores to have the same task that they can work on simultaneously. CPU cores also require higher memory storage than GPU cores. On the other hand, GPUs contain a larger volume of weaker cores that perform specialized calculations in a parallel manner. They work on more specialized tasks while the CPU works on more general tasks. They also don’t require as much memory as CPU cores do.
Software Threads#
GPUs and CPUs contain a different number of cores, which directly affects each microprocessor’s performance. CPUs only have cores that number by the dozen, and so they can only execute a limited number of commands at a time. Granted, CPUs have multi-threading capabilities, but these can only go so far as to speed up the switch between pre-set tasks. On the other hand, GPUs are equipped with hundreds of cores to handle thousands of software threads simultaneously. Moreover, GPUs are more programmable today than they were in the past. This could mean two things. The first is that you can improve speed and performance by using overclocking software. The other is that you can program the GPU to work on specialized tasks.
Processing Method#
Engineers designed CPUs more for serial processing and GPUs for parallel processing. Serial processing is when a task is wholly assigned to a single core, and the tasks are completed on a sequential basis. GPUs, on the other hand, divide the task amongst several cores that execute the sub-tasks simultaneously. The difference is that a single core in a CPU has a higher workload than a single core in a GPU.
Latency#
The number of cores also contributes to how fast the processors can complete computing tasks. CPUs, for example, take longer than GPUs to accomplish tasks for this exact reason. The limited number of cores that CPUs have limit the concurrency of tasks that they can accomplish. At the opposite end are GPUs with many cores that can process large volumes of data simultaneously and non-sequentially. This distributive nature of GPU processing allows it to save a lot of time. Another reason why GPUs have better speeds has to do with the fact that they have VRAM memories. VRAM stands for video random access memory. It’s a special type of RAM that provides access to the information on the CPU. The availability of that information on the VRAM saves the CPU on time normally used to transfer the data to the GPU.
Efficiency#
The root of the matter relates to the architectural differences in the cores of CPUs and GPUs. GPUs consume less power than CPUs in general. GPU cores are more resource-efficient, which means they perform more work for every energy unit they receive than CPU cores. The power efficiency offered by GPUs makes it ideal for cloud computing and big data analytics. A few factors contribute to GPU’s power efficiency. The first relates to data latency, the time it takes to travel from the main memory and down into the processor. Engineers designed CPUs to have smaller bodies with tightly packed transistors. On the other hand, you have CPUs with larger cores that are farther apart. This increases the speed of processing for GPUs compared to CPUs. Additionally, GPU cores have the uncanny ability to transfer data into all cores simultaneously. On the other hand, CPUs have to check every core in the system before transferring data and executing commands.
Host Codes#
CPUs typically work with the very basic units of coded language known as “machine codes” or “native codes.” Machine codes contain a series of instructions that the CPU must relay to other components. Any given software might contain millions of machine codes strung together. The CPU converts the codes into instructions which it then relays to the appropriate components. For example, software as common as Firefox comprises 21 million lines of code. On the other hand, GPUs can only run a special type of code called “shaders.” Shaders are additional bits of code that help to improve the graphics quality of games and other programs. Enhancing graphics post-production requires a lot of complicated math to execute. Luckily, GPU manufacturers have designed GPUs to handle exactly these types of tasks. The presence of a GPU frees up the CPU so it can handle more important tasks. These might include rendering, compression, or transcoding tasks, to name a few. Nonetheless, the CPU is still responsible for sending all the information to the GPU for processing.
GPU vs CPU: A Paradigm Shift#
GPUs and CPUs are tied to the hip when it comes to their functions. They complement each other more often than they compete with each other. But as recent developments will prove, they are likely to evolve into integrated forms or merge with other components into a single chip. Here are some examples of hybrid microchips that may blur the GPU vs CPU lines in the near future.
Accelerated Processing Units (APUs)#
During the past decade, various efforts to integrate CPUs and GPUs into a single device have come about. A few manufacturers who succeeded in this effort called their products accelerated processing units (APUs). APUs allow you to enjoy the same benefits that a typical GPU and CPU setup does on your computer. The GPU and CPU combination will speed up your applications, improve the quality of your graphics, as well as decrease power consumption over time. Manufacturers combine the two components into a single circuit or “die” secured by the semiconductor material. Keeping the two components tightly packed next to each other boosts data frame rates and reduces power consumption. APU designs also come with cost savings for both manufacturers and consumers. Not to mention, the compact design also means that you have more room for other types of hardware.
System-on-a-Chip (SoC)#
If combining two processors into a die could bring performance enhancements, then it follows that adding more systems onto a single chip would have the same effect. If you could add two components and turn them into a single device, then perhaps it’s also possible to merge multiple components into a single device. A System-on-a-Chip (SoC) is any device that fits into this description. An SoC is an integrated component that combines a variety of essential components into single hardware. Like APUs, SoCs incorporate GPUs and CPUs into a single chip. Except that they also include memory and secondary storage, plus other things that the manufacturer might want to add. The benefits of SoCs may be even greater than plain APUs. The benefits of APUs, including more efficient power consumption and heat generation, plus increased performance, also hold for SoCs. SoCs also have a smaller footprint since all the components are on the same chop and internally connected. A single SoC chip is a testament against the GPU vs CPU comparison, and it’s more cost-efficient than purchasing individual components. Also, read about the original Apple M1 Mac, the first Apple-designed System on a Chip (SoC) device with both CPU and GPU components. If you need a broader perspective on how the M1 compares to other brands, read this article comparing the Apple M1 Chip with the Intel Core i7 CPU.
GPU vs CPU: The End of Moore’s Law#
Despite the obvious differences within the GPU vs CPU discussion, both microprocessors remain the heart and soul of our electronic devices. Without these processors, our computers would be nothing more than hardware put together. Each has its advantages that come in handy for keeping up with performance standards on computer systems. But what changes can we expect for these crucial components? Let’s find out.
Moore’s Law#
Computer hardware is always evolving, with components becoming smaller and more powerful. This is the foundation for Moore’s law. Moore’s law is a scientific observation that microprocessors are constantly improving, with a new release every two years. As time goes on, microprocessors contain more transistors, and they also arrive in increasingly small packages. The historical data seems to support this prediction, with new and more powerful transistors coming out every two years or so. Intel invented the very first microprocessor back in 1971, and it had 2,300 transistors. Nowadays, most processors contain billions of transistors that may directly or indirectly control other components. Reducing the size of microchips is another factor in its development to make it portable. Microprocessors in the 2000s were about 90nm in size, but today’s microprocessors are anywhere between 5 to 14 nm in size.
A Self-Fulfilling Prophecy#
Moore’s law was a kind of self-fulfilling prophecy. Over the past few decades, scientists have managed to scale down the size of microprocessors and pack in as many transistors into them as possible. But this improvement cannot last forever, and at some point, we will hit a wall with how fast and how small our transistors can be. Let’s take the eight-core Apple M1 Pro chip as an example. It contains 33.7 billion transistors, more than twice the original M1 chip. The manufacturers compressed the transistors into a silicone packet about the size of a quarter of a post-it note. Anything lower than this level of compaction and the laws of physics will start to go against you. That is, electrons will begin to go haywire and end up in places where they shouldn’t be. In addition that, it will be impossible to fit the transistors into silicon wafers, which control their electronic properties.
Moore’s Law and Processing Speeds#
But what does Moore’s law have to do with GPUs vs. CPUs? The answer is, well, everything. Remember when we mentioned earlier that both microprocessors are made out of arithmetic logic units (ALUs)? In truth, the ALUs contain thousands of transistors (per unit). These transistors act as electronic switches that control the speed and function of GPUs and CPUs. And of course, they are the topic of discussion under Moore’s law. These transistors are responsible for speed improvements in CPUs and even GPUs. Once the validity of Moore’s law expires, there would no longer be any room to move when it comes to microprocessor speeds. In fact, the production of microchips has become so expensive and time-consuming that many manufacturers have abandoned the effort. As such, only a handful of manufacturers are continuing the effort to develop new microchips. The experts say that we have anywhere between eight to ten more cycles before Moore’s law finally expires, which is good news. Come to think of it, though, that last microprocessor will still bring us mind-numbingly fast speeds. Just imagine the current speed of the world’s fastest processor, the AMD Ryzen 1950X’s 3.4 GHz speed, and then squaring it by a factor of six. So while we will inevitably hit that wall in the future, it likely won’t be as bad as expected.
Impact of Moore’s Law on CPUs and GPUs#
With Moore’s law bound to happen down the line, many people wonder if this is the end for GPUs and CPUs. The answer is probably not. CPUs and GPUs both perform essential tasks on computers, and as far as we’re concerned, there are no substitutes for these microprocessors and the way they work. The more likely scenario is that GPUs and CPUs will continue to improve over time, up until they reach the point where it’s physically impossible to fit transistors into a smaller space. Even when we reach that point, the more likely scenario is that silicon transistors will be replaced, not the processors. Scientists are already looking at alternatives, which include circuit memory capabilities called “memristors.” It’s also unlikely that the CPU will replace the GPU. Because while GPUs have significantly more cores than CPUs, their cores are far slower. Additionally, manufacturers designed GPUs to handle compute-heavy tasks instead of generic computing tasks. They rely entirely on the CPU for information and instructions. Therefore both microprocessors are likely to carry on with their respective roles.
Final Thoughts#
CPUs and GPUs are comparable to a Swiss knife and combat knife, respectively. The Swiss army knife is useful for many different things, such as opening cans or cutting small pieces of firewood in the forest. But you wouldn’t want to use the Swiss army knife once an enemy comes along; you need the combat knife for that. Following this analogy, CPUs cater to various generic computing tasks, while the GPUs are more focused on specialized, compute-heavy tasks. Nevertheless, both microprocessors remain vital to the performance of computer systems, so it’s not a matter of choice. Both components are equally necessary for excellent performance, and the system benefits the most when both components are present. That settles our discussion on GPU vs CPU
В чем разница между CPU и GPU
Центральные и графические процессоры похожи, они сделаны из миллионов транзисторов и обрабатывают тысячи операций в секунду. Несмотря на их сходства, изначально их создавали для решения разных задач. В статье разбираемся, чем отличается CPU и GPU и какие задачи решают.
Что такое CPU и GPU
Центральный процессор (CPU ) — это центральный узел, который управляет всеми процессами в компьютере. Например, выполняет арифметические и логические операции с данными, передает результаты на внешние устройства, хранит результаты выполненных операций. С CPU работают любые гаджеты, будь то планшет или смартфон.
Графический процессор (GPU) — это микропроцессор, который ускоряет запуск графики, которая отображается на экране. Раньше его использовали в основном для рендеринга изображений, ЗD-моделирования. Сейчас же GPU применяют для высокопроизводительных вычислений, машинного обучения и научных исследований.
Различия между GPU и CPU
Количество ядер
У CPU обычно есть 4-8 ядер, которые работают независимо друг от друга. Архитектура CPU обычно основана на принципе меньшего числа, но более мощных ядер обработки информации. В GPU ядер значительно больше — от 100 до 1000, но они меньше по мощности, чем у CPU.
Отличия в количестве ядер у центрального и графического процессоров
Способ обработки данных
CPU — это универсальный процессор, он выполняет последовательные задачи. Например, вычисления, управление операционной системой и выполнение программ.
Как работает CPU
В свою очередь, графический процессор обрабатывает данные параллельно, что позволяет ему выполнять интенсивные задачи, такие как рендеринг графики, 3D-моделирование, машинное обучение и HCI.
Как работает GPU
Использование памяти устройства
Для центрального процессора кэш-память — это ключевой параметр, поэтому он занимает большой объем памяти устройства. Графическому процессору не нужна кэш-память большого размера. Например, для рендеринга изображений достаточно 128-256 кБ.
Скорость вычислений
Тактовая частота — один из главных технических показателей любого процессора, который измеряют в герцах. Чем этот показатель выше, тем больше вычислений за единицу времени успеет сделать оборудование. В отличие от CPU, приложения на графическом процессоре работают быстрее, без скачков.
Количество потоков
Центральный процессор поддерживает до двух потоков вычислений на одно ядро, а графический — несколько тысяч потоков на каждый мультипроцессор, которых в чипе несколько штук. Чтобы переключиться с одного потока на другой, CPU необходимы сотни тактов, а GPU переключает несколько потоков за один такт.
Задачи CPU и GPU
Когда используют CPU:
- работа с общими приложениями, такими как обработка текстов, таблиц или других офисных документов;
- обработка больших объемов данных, но эти данные не являются графикой;
- выполнение нескольких задач последовательно.
Когда лучше подойдет GPU:
- работа с графикой, такой как 3D-моделирование, визуализация данных, обработка изображений или видео;
- машинное обучение или научные расчеты, которые требуют вычислительной мощности;
- использование параллельных алгоритмов.
Когда CPU и GPU работают вместе
CPU и GPU могут работать вместе, чтобы увеличить пропускную способность данных и одновременных вычислений. GPU может дополнить архитектуру CPU, чтобы выполнять повторяющиеся вычисления параллельно, в то время как остальные действия выполняются последовательно на CPU. Плюс каждый из процессоров занимается своими непосредственными задачами: GPU — выполняет сложные вычисления, а CPU координирует широкий спектр действий.
В чем отличия процессоров для компьютеров и серверов
Компьютерные процессоры отличаются по архитектуре от серверных. К CPU на серверах выдвигают особые требования по надежности и безотказности. Эти процессоры ориентированы на круглосуточную работу и высокие нагрузки.
GPU для серверов тоже заточены под нагрузку в режиме non-stop. Даже если внешне процессоры для компьютера и сервера будут похожи, их технические характеристики будут отличаться. Nvidia предлагает для серверного оборудования специальные линейки GPU с названием TESLA и QUADRO. У AMD для серверов есть FirePro или Radeon & Vega.
Аренда облачных серверов с GPU
Сейчас аренда облачных серверов с GPU — одна из популярных услуг для бизнеса. Компании получают вычислительные мощности для решения разнообразных задач без необходимости инвестиций в оборудование и долгого ожидания поставок. В облаке можно использовать ресурсы в зависимости от потребностей проекта.