This era is a full combination of technologies, software, hardware. All the software, hardware, and technologies are a result of different types of processing units and their cores. Such a revolutionary core is the CUDA core.
CUDA core comes with the making of power imagination into reality. All the unreal graphics designs now come into reality because of this core. The artificial intelligence has got a shape and works for mankind because of this remarkable core. So have a look toward this core below.
What Is CUDA Core?
CUDA stands for ‘’Compute Unified Device Architecture’’. This extension doesn’t tell us about the working procedure of CUDA or what CUDA is and does. Again, core stands for the central part of something, I mean, very important. So you can say CUDA is something that is an essential part of some particular computing. The CUDA core is a software layer that uses the parallel computing process.
NVIDIA is a technology company that designs GPUs( Graphics Processing Units) and they have created CUDA as a software platform that pairs with their GPU hardware making it easier for developers to build software that accelerates computations using the parallel processing power of NVIDIA GPUs.
NVIDIA GPU is the hardware that enables parallel computations while CUDA is the software layer that provides an API(Application Programming Interface) for developers. As a result, you might have guessed that NVIDIA GPU is required to use CUDA.
Once you have an NVIDIA GPU, CUDA can be downloaded and installed from NVIDIA’s website for free. Developers use CUDA by downloading the CUDA toolkit, in with the CUDA toolkit come specialized libraries like cuDNN.
In a word, CUDA is a core that is completed through the GPU. Where CUDA is a software platform to do parallel computing, there GPU is a hardware platform to accomplish that computing. As CUDA is a software, it uses C, C++, Fortran, and Python according to its need.
How Does CUDA core Work?
To know CUDA, we have to know the working procedure of Graphics Processing Units(GPUs) because GPU consists of hundreds or thousands of CUDA cores. So a GPU is a processor that is a combination of cores, is good at handling specialized computations, known as parallel computing.
The type of computation most suitable for GPU is a computation that can be done in parallel, CUDA core brings us to parallel Computing. Parallel computing is a type of computation whereby a particular computation is broken into independent smaller computations that can be carried out simultaneously. Those independent smaller computations are done by each CUDA core.
CUDA core is a program based platform that does some specific work through programming. When work comes to be executed, then firstly that work is divided into some individual parts. After that, according to programming each smaller part of work is executed by each CUDA core that embedded on the GPU.
As all the CUDA cores are busy in doing individual tasks, no time is being wasted. CUDA core at last compels GPU to process all the individual tasks and get the desired output.
CUDA core is busy only on completing graphics work. It makes the graphics look more realistic as it provides high-resolution graphics. So where is graphics related works, there stands CUDA core with its parallel computation.
List of CUDA Core
You will get the general information from the list of CUDA core that is based on official specifications. There are almost 10-12 or more versions of CUDA cores from the beginning to now. They are:
1. CUDA SDK 1.0
Here SDK stands for software development kit. So this 1.0 stands for the computing capability. This version of CUDA can support the computing capability from 1 to 1.1(Tesla). This type of core is used in G80 GPUs.
2. CUDA SDK 1.1
This CUDA core is used for the computing capability of 1 to 1.1+x (Tesla). This one is applicable for G84, G86, G92, G94, G96, G98 GPUs.
3. CUDA SDK 2.0
This one is used to support for compute capability 1.0– 1.1+ x (Tesla). This goes with GF 100 and GF 110 GPUs.
4. CUDA SDK 2.1 – 2.3.1
Compute capability ( 1.0 – 1.3 )Tesla for this version. These are applicable for GF104, GF106 GF108, GF114, GF116, GF117, GF119 GPUs.
5. CUDA SDK 3.0 – 3.1
The compute capability is 1.0 – 2.0 (Tesla). GK104, GK106, GK107, these GPUs can use these CUDA cores.
6. CUDA SDK 3.2
This one is for compute capability 1.0 – 2.1 (Tesla). GK20A GPUs is used this core and the microarchitecture is named after Kepler.
7. CUDA SDK 4.0 – 4.2
It supports for compute capability 1.0 – 2.1+x (Tesla, Fermi).
8. CUDA SDK 5.0 – 5.5
The computing capability of this CUDA core is 1.0 – 3.5 (Tesla, Fermi, Kepler). The microarchitecture of these is named after Maxwell. GM107, GM108, GM200, GM204, GM206, GM20B, all these GPUs consists of this CUDA cores.
9. CUDA SDK 6.0
The computing capability is 1.0 – 3.5 (Tesla, Fermi, Kepler). The microarchitecture of these is named after Pascal. GP102, GP104, GP106, GP107, GP108, and GP10B consist of this SDK of CUDA core.
10. CUDA SDK 6.5
support for compute capability 1.1 – 5.x (Tesla, Fermi, Kepler, Maxwell). The microarchitecture of these is named after Pascal.
11. CUDA SDK 7.0 – 7.5
The computing capability is 2.0 – 5.x (Fermi, Kepler, Maxwell). The microarchitecture of these is named after Volta and Turing. GV100, GV10B, TU102, TU104, TU106, TU116, TU117, all these GPUs use this core.
12. CUDA SDK 8.0
support for compute capability 2.0 – 6.x (Fermi, Kepler, Maxwell, Pascal). The microarchitecture of these is named after Ampere.
13. CUDA SDK
9.0 – 9.2 support for compute capability 3.0 – 7.2 (Kepler, Maxwell, Pascal, Volta).
14. CUDA SDK 10.0 – 10.2
This CUDA SDK supports for compute capability 3.0 – 7.5 (Kepler, Maxwell, Pascal, Volta, Turing).
15. CUDA SDK 11.0
This one is under construction, so no specification has been achieved till now.
CUDA core VS CPU core
We know the CUDA core is embedded in GPUs. So it can be called GPUs core. There is a great deal of difference between CUDA core and CPU core. CUDA core works for GPUs, but CPU core works for CPUs.
GPUs work in parallel processing, so CUDA cores also execute in the parallel pipeline. Each CUDA core works for the same code that other does at the same time in parallel, so no need to wait for one core to accomplish its job. Rather all the CUDA cores can do their job together but independently. No extra time is needed and wasted in the CUDA core.
But in CPU cores, they prefer series computation. So all the cores need to wait for the completion of the job of the previous core. Each CPU core contains registers, cache memory, so each of them does their work in series one after another. That’s why sometimes the CPU core needs more time.
As a metaphor, if CPU can be considered as a hypercar, then CUDA core can be called a huge dump truck. If a simple operation is needed to execute than it is easy to complete. Then, the CPU core needs less time. But CUDA core works with mainly complex operations like graphics problems. So it needs more time.
As CPU core is now generalized core, so it normally works faster than CUDA core. Because the CUDA core has to maintain many operations at a time, but the CPU core doesn’t need it.
Specifically, CPU cores are general purpose unit for your PC or laptop, but CUDA core is a specific purpose, like graphics design, AI, unit.
CUDA core VS Tensor core
CUDA core is a stream processor, but a tensor core a stream processor in the simplest form. CUDA core works for different aspects regarding graphical problems and artificial intelligence. In the case of tensor core, it works only for one aspect of CUDA core which lying in matrix multiplication and addition.
Tensor Core is a working core of mixed precision training that is FP16 matrix multiply addition. It takes in 16-bit floating point, then does this matrix multiply, then accumulates everything in 32-bit and 32-bit precision accumulation tends to matters for convergence of network to make a mixed-precision network. That’s how a tensor core works. In a word, tensor work works way faster than CUDA core in deep learning.
Tensor core is not best for all parallel computation rather better for only one computation that is related to square matrices. So it is massively faster than CUDA core as CUDA is applicable for all types of parallel computation.
Tensor cores are the specialist or like the genius in a specific field, but CUDA core is a lecture of the field.
Conclusion
CUDA core is a demand of time nowadays. The whole world now works and lives on graphics designs and artificial intelligence. So it is high time to know about CUDA core and usages. Here you will have all the information about the CUDA core that you search for. So welcome the era of knowledge of CUDA core. I hope that you are going to get a good time. Thank you.
Leave a Reply