The main differences are the number of cores, processor speed, available power, and memory. Integrated GPUs are part of another system component, usually the CPU nowadays. That limits the number of cores the GPU part can have, as well as their speed and power. They also have to use part of the system memory for its own tasks. Nonetheless, they're more energy efficient and more than powerful enough for most common tasks. Discrete GPUs have their own dedicated processor and high-speed memory. Higher end GPUs typically require more power, sometimes even more than the rest of the system.
A good follow up video would be a similar short summary explainer on SoCs. And maybe one on the difference between the other key components like Media Engines, NPUs, onboard vs separate RAM etc. Thanks. 🙏🏽
I think pretty much everyone knows the difference between gpu and cpu, the most useful information here would be the "why" the gpu cannot be used as a CPU.
If possible, explain why a GPU can not replace a CPU. Great vid but my old vic64 brain (yeah im old) dont get this. Anyway, cheers from a new subscriber
When you stress the system by forcing a dos game to play the intro at very high cpu cycles, you also frce the system to turn on secret accelerating features that remain on. One is the sound acceleratig of sound device, it offlads the cpu from decompressing audio. The ther is the floating point unit, that is off by default, when it is on, some games will become harder. The Intel has an autopilot for car games, that is also off by defaut. With the GPU the secret seems to be the diamnd videos I am experimenting with, many games show diamnds as reward, and hard to get. Diamnd stresses the vga card, it's so compex to draw by the video chip. Also the tuning will consume the battery faster.
So let me ask a stupid question, why can't they combine the two and get the best of both worlds?Will doing so negate some functions rendering it useless? Or .. lol just curious 🤔 😊
CPU has few cores which is why parallel processing is hard on them but each individual core is extremely powerful. GPU has thousands of cores making them good at parallel processing but each individual core is comparatively not very powerful. You can make a Processor which has multiple cores like GPU and each core being powerful like CPU but that would cost huge amount of money
The question is stupid. You can either get high performance in non-parallel tasks and low latency or high performance in parallel tasks and high latency, if you extend the normal core with GPU features, it will instantly become large, power hungry and slow for normal tasks, that is why modern CPUs for desktops have bult-in separate GPU cores. Also, the problem is that you would need a lot of memory channels of memory to make use of that - memory for GPUs is very slow, but has a lot of internal channels.
What do you think will happen if both the heat of cpu ang separate gpu combined with just one cooling fan and heatsink? Plus greater voltage needs resulting in more heat?
In neural networks it's all about matrix multiplications and also we want to pass multiple inputs through the network (which also perform a number of matrix multiplication) with GPU we can perform all the passes of different inputs to the neural network parallel instead of just doing one input a time which can speed up the computations
Still not clear for me: what component does the GPU miss so that it cannot replace a CPU? Ah, just checked with perplexity ai, the instruction set that a GPU accepts is too poor to make a GPU a replacement for a CPU.
gpu is like PhD holder while the CPU is the jack of all trades. look at the name GPU graphics computing CPU is Central processing unit GPU can outperform cpu computing power in singular task like in graphics computing or any other given task, while the cpu can't compute as fast as the gpu cpu can compute task simultaneously.
Most of things in this video is BS. GPU can replace CPU, but will work many times slower for most of tasks while taking more power. Only on highly parallel tasks it is efficient and fast. Also, it is missing a lot of features to control hardware, like proper interrupts management etc.
@@avalagum7957 It is optimized to consume minimum amount of energy and perform multiple calculations per cycle, but the calculations take much longer time to finish, up to 100 times slower. All those parallel operations go to waste if you don't need to perform the exact same operation on multiple entries. Also, its memory is way slower than that of CPU (albeit CPU memory is also not very fast, it merely got ~30% faster over the last 20 years, but got a lot of internal channels), but contains a lot of internal channels, so it is efficient in processing large amounts of data, which do not need high performance per each dataset.
You are wrong about many things. GPUs, the modern ones, aren't actually good in performing operations for each single pixels. They are far behind the CPU on that, but can more efficiently work with large groups of pixels. No, the modern high-end GPUs have 32-64 cores (top ones like 4090 have 128 cores). The marketing cores is a lie. No, the threads is a lie, the actual number is hundreds times lower. Those fake threads are parallel execution units, they are not threads, they are same code which works with a very large array. Each single core can run 1 or 2 actual threads, therefore the number of threads for high-end GPUs is usually limited to 128 or so. Only because of repetitive operations GPUs are faster in some tasks, they are generally much slower than a crappy processor.
A thread is simply a logical unit of sequence of instructions that gets mapped to each core by a scheduler (modern systems usually use both hardware and software scheduling these days) whether that's on a GPU or CPU for execution. Primary difference between GPU and CPU threads is that GPUs usually execute the same "copy" of each instruction across all threads being executed on the device (single-instruction multiple data aka SIMD), whereas CPUs can have all the different threads very easily execute all kinds of different instructions simultaneously (multiple-instruction multiple data aka MIMD). In addition more than one thread can be mapped to each core whether it be a CPU or GPU and in the case that happens simultaneous multi threading (SMT) hardware technology is used to execute all the threads at once.
Depends on your definition of "have": if Nvidia makes a GPU driver that supports the CPU in question then technically one can combine an Nvidia GPU and that CPU into the same computer to run CUDA stuff, but executing GPU CUDA code on the CPU is something that Nvidia probably doesn't want people to do since Nvidia likely wants to keep making money on their GPU sales so executing GPU CUDA code on CPU will likely not be a thing anytime soon
Does this means that gpu is a better version of cpu only or we can say faster version of cpu which can do many calculation at a time, then why dont we use two gpus instead of a cpu and a gpu?
GPU only performs better if the task being performed can be parallelized, but majority of the tasks can't be or don't need parallel computation So they would be slower on GPU. Main Power GPU gives us is parallelization if can't exploit it in a task the overhead will make it even slower than CPU.
No, you didn't explain nor do you understand anything, all you did was read a wikipedia page, gpus dont have "more cores" thats a marketing lie, what they have are very wide simd lanes, cpus have it too but they are smaller in exchange for bigger cache, higher frequency and less heat.
Honestly this video despite oversimplifying a bit (due to time constraints likely) isn't entirely incorrect. There are in fact more cores on average in a GPU it's just the architecture of each GPU core is completely different than a CPU core; companies like Nvidia and AMD aren't lying when they talk about their CUDA core/stream processor counts it's just that each core in a GPU serves somewhat different purposes than each core in a CPU is all. Also what you're saying about cache isn't really correct AMD RDNA 2+ cards have a pretty big pool of last level infinity cache that contributes a significant amount to the overall GPU performance
@@mattslams-windows7918 The problem is, as a graphics programmer, if i were to learn it again, this video tells me nothing, its literally a presentation a high school student would do for homework "cpus do serial gpus do parallel", the funny thing is that the main differences are literally shown in the first image he showed, where it shows the number of ALUs (the simd lanes, they also have FPUs) of each, yet he didnt explain that because he has no clue what those images mean. now gpus do have slightly more cores than their cpu counterpart, but its usually 1.5x-2x higher, not thousands of cores, if you want the correct terminologies, a cpu core is equivalent to streaming multiprocessor in nvidia , compute unit in amd(funny enough amd also refers to them as cores on their integrated graphics specifications), and core in apple, a cpu thread is warp in nvidia and wavefront in amd, a cpu simd lane is a cuda core in nvidia and stream processor in amd, now for the cache thing you mentioned, you are probably using a gaming pc or console for the comparison, they will usually have 4-8 core cpus with 16-32 core gpus, in games single core performance matters more (usually because most gamedevs dont know how to multithread haha), if you want more less of a comparison take the ryzen 9 5900x 12 core and the rx 6500 16 cores, roughly similar power consumption, L3 cache is 64MB on the cpu and 16MB on the gpu, L2 cache is 6MB on the cpu and 1MB on the gpu, L1 cache is 768KB and 128KB on the gpu, now if you get a gpu with higher core counts, you will notice that L3 cache increases a lot but L1 cache stays the same, this is because L3 cache is a shared memory pool for all of the cores within the gpu and cpu, while L2 and L1 cache are local to the core. anyways that was a long reply, hopefully that answered your questions xD.
![HIWWHEE | OHM [Official MV]](http://i.ytimg.com/vi/OJJLtTLqLag/mqdefault.jpg)
Can you make a similar video for GPU vs Integrated GPU? Is there any difference in their architectures?
The main differences are the number of cores, processor speed, available power, and memory.
Integrated GPUs are part of another system component, usually the CPU nowadays. That limits the number of cores the GPU part can have, as well as their speed and power. They also have to use part of the system memory for its own tasks. Nonetheless, they're more energy efficient and more than powerful enough for most common tasks.
Discrete GPUs have their own dedicated processor and high-speed memory. Higher end GPUs typically require more power, sometimes even more than the rest of the system.
Dont forget compute unit.
finally an explanation that I completely understand and not trying to sell me on anything!
Simple yet short to the point, instant sub thanks mate.
That was the goal glad it was helpful!
Beautifully done. The best explanation I came across. Understood the core concepts you explained. Again, beautifully executed
Thanks for the kind words!
Very clearly explanations!
Briljant video! I started my IT Forensics study last week and will share this with the other students in my class!
Glad it was helpful!
With the dawn of AI PCs, which always have CPU, GPU and NPU.
Can you make a similar video on differences between GPU and NPU ?
Its' on the list!
A good follow up video would be a similar short summary explainer on SoCs. And maybe one on the difference between the other key components like Media Engines, NPUs, onboard vs separate RAM etc. Thanks. 🙏🏽
Great idea, I'll put it on the list!
Very powerful and concise explanation. Keep up the good work.
Thank you!!
Thank you... Simple and short....
Thank you!
Very clear, thank you
I subscribed to your channel to see more videos like this.
Thank you!!
Beautifully explained. Thank you
Intel Core2 architecture can execute up to 4 integer instructions paralell with each single core.
Excellent explanation. Thanks!.
Thank you!
Great!
Holy Cow I’m ready to be a computer engineer. 👍🏾👌🏾🤓
2:34 lol, *core* differences
I think pretty much everyone knows the difference between gpu and cpu, the most useful information here would be the "why" the gpu cannot be used as a CPU.
If possible, explain why a GPU can not replace a CPU. Great vid but my old vic64 brain (yeah im old) dont get this. Anyway, cheers from a new subscriber
helpful. thank you!
Thanks!
Excellent!
Thank you!
When you stress the system by forcing a dos game to play the intro at very high cpu cycles, you also frce the system to turn on secret accelerating features that remain on. One is the sound acceleratig of sound device, it offlads the cpu from decompressing audio. The ther is the floating point unit, that is off by default, when it is on, some games will become harder. The Intel has an autopilot for car games, that is also off by defaut.
With the GPU the secret seems to be the diamnd videos I am experimenting with, many games show diamnds as reward, and hard to get.
Diamnd stresses the vga card, it's so compex to draw by the video chip. Also the tuning will consume the battery faster.
So let me ask a stupid question, why can't they combine the two and get the best of both worlds?Will doing so negate some functions rendering it useless? Or .. lol just curious 🤔 😊
CPU has few cores which is why parallel processing is hard on them but each individual core is extremely powerful.
GPU has thousands of cores making them good at parallel processing but each individual core is comparatively not very powerful.
You can make a Processor which has multiple cores like GPU and each core being powerful like CPU but that would cost huge amount of money
The question is stupid. You can either get high performance in non-parallel tasks and low latency or high performance in parallel tasks and high latency, if you extend the normal core with GPU features, it will instantly become large, power hungry and slow for normal tasks, that is why modern CPUs for desktops have bult-in separate GPU cores. Also, the problem is that you would need a lot of memory channels of memory to make use of that - memory for GPUs is very slow, but has a lot of internal channels.
Can you combine a school bus with a Formula One car?
There, you have your answer.
APUs do that.
What do you think will happen if both the heat of cpu ang separate gpu combined with just one cooling fan and heatsink? Plus greater voltage needs resulting in more heat?
What is it about AI that requires intense parallel computation?
In neural networks it's all about matrix multiplications and also we want to pass multiple inputs through the network (which also perform a number of matrix multiplication) with GPU we can perform all the passes of different inputs to the neural network parallel instead of just doing one input a time which can speed up the computations
Still not clear for me: what component does the GPU miss so that it cannot replace a CPU?
Ah, just checked with perplexity ai, the instruction set that a GPU accepts is too poor to make a GPU a replacement for a CPU.
gpu is like PhD holder while the CPU is the jack of all trades.
look at the name GPU graphics computing CPU is Central processing unit
GPU can outperform cpu computing power in singular task like in graphics computing or any other given task, while the cpu can't compute as fast as the gpu cpu can compute task simultaneously.
Most of things in this video is BS. GPU can replace CPU, but will work many times slower for most of tasks while taking more power. Only on highly parallel tasks it is efficient and fast. Also, it is missing a lot of features to control hardware, like proper interrupts management etc.
@@trevoro.9731 why does GPU is slower than cpu for most of tasks?
@@avalagum7957 It is optimized to consume minimum amount of energy and perform multiple calculations per cycle, but the calculations take much longer time to finish, up to 100 times slower. All those parallel operations go to waste if you don't need to perform the exact same operation on multiple entries.
Also, its memory is way slower than that of CPU (albeit CPU memory is also not very fast, it merely got ~30% faster over the last 20 years, but got a lot of internal channels), but contains a lot of internal channels, so it is efficient in processing large amounts of data, which do not need high performance per each dataset.
You are wrong about many things.
GPUs, the modern ones, aren't actually good in performing operations for each single pixels. They are far behind the CPU on that, but can more efficiently work with large groups of pixels.
No, the modern high-end GPUs have 32-64 cores (top ones like 4090 have 128 cores). The marketing cores is a lie.
No, the threads is a lie, the actual number is hundreds times lower. Those fake threads are parallel execution units, they are not threads, they are same code which works with a very large array. Each single core can run 1 or 2 actual threads, therefore the number of threads for high-end GPUs is usually limited to 128 or so.
Only because of repetitive operations GPUs are faster in some tasks, they are generally much slower than a crappy processor.
What about the Threads?
A thread is simply a logical unit of sequence of instructions that gets mapped to each core by a scheduler (modern systems usually use both hardware and software scheduling these days) whether that's on a GPU or CPU for execution. Primary difference between GPU and CPU threads is that GPUs usually execute the same "copy" of each instruction across all threads being executed on the device (single-instruction multiple data aka SIMD), whereas CPUs can have all the different threads very easily execute all kinds of different instructions simultaneously (multiple-instruction multiple data aka MIMD). In addition more than one thread can be mapped to each core whether it be a CPU or GPU and in the case that happens simultaneous multi threading (SMT) hardware technology is used to execute all the threads at once.
can cpu have something like cuda?
Depends on your definition of "have": if Nvidia makes a GPU driver that supports the CPU in question then technically one can combine an Nvidia GPU and that CPU into the same computer to run CUDA stuff, but executing GPU CUDA code on the CPU is something that Nvidia probably doesn't want people to do since Nvidia likely wants to keep making money on their GPU sales so executing GPU CUDA code on CPU will likely not be a thing anytime soon
Moore Maria Williams Edward Johnson Richard
Gonzalez Frank Smith Christopher Thomas Anna
Does this means that gpu is a better version of cpu only or we can say faster version of cpu which can do many calculation at a time, then why dont we use two gpus instead of a cpu and a gpu?
GPU only performs better if the task being performed can be parallelized, but majority of the tasks can't be or don't need parallel computation So they would be slower on GPU. Main Power GPU gives us is parallelization if can't exploit it in a task the overhead will make it even slower than CPU.
No, GPU is like a factory with 100 workers. CPU is like a medical practice with 4 doctors. Neither can do each other's job.
Since each core is similar to CPU can we say that it has multiple CPU units?
No, you didn't explain nor do you understand anything, all you did was read a wikipedia page, gpus dont have "more cores" thats a marketing lie, what they have are very wide simd lanes, cpus have it too but they are smaller in exchange for bigger cache, higher frequency and less heat.
Honestly this video despite oversimplifying a bit (due to time constraints likely) isn't entirely incorrect. There are in fact more cores on average in a GPU it's just the architecture of each GPU core is completely different than a CPU core; companies like Nvidia and AMD aren't lying when they talk about their CUDA core/stream processor counts it's just that each core in a GPU serves somewhat different purposes than each core in a CPU is all. Also what you're saying about cache isn't really correct AMD RDNA 2+ cards have a pretty big pool of last level infinity cache that contributes a significant amount to the overall GPU performance
@@mattslams-windows7918 The problem is, as a graphics programmer, if i were to learn it again, this video tells me nothing, its literally a presentation a high school student would do for homework "cpus do serial gpus do parallel", the funny thing is that the main differences are literally shown in the first image he showed, where it shows the number of ALUs (the simd lanes, they also have FPUs) of each, yet he didnt explain that because he has no clue what those images mean.
now gpus do have slightly more cores than their cpu counterpart, but its usually 1.5x-2x higher, not thousands of cores,
if you want the correct terminologies,
a cpu core is equivalent to streaming multiprocessor in nvidia , compute unit in amd(funny enough amd also refers to them as cores on their integrated graphics specifications), and core in apple,
a cpu thread is warp in nvidia and wavefront in amd,
a cpu simd lane is a cuda core in nvidia and stream processor in amd,
now for the cache thing you mentioned, you are probably using a gaming pc or console for the comparison, they will usually have 4-8 core cpus with 16-32 core gpus, in games single core performance matters more (usually because most gamedevs dont know how to multithread haha),
if you want more less of a comparison take the ryzen 9 5900x 12 core and the rx 6500 16 cores, roughly similar power consumption, L3 cache is 64MB on the cpu and 16MB on the gpu, L2 cache is 6MB on the cpu and 1MB on the gpu, L1 cache is 768KB and 128KB on the gpu, now if you get a gpu with higher core counts, you will notice that L3 cache increases a lot but L1 cache stays the same, this is because L3 cache is a shared memory pool for all of the cores within the gpu and cpu, while L2 and L1 cache are local to the core.
anyways that was a long reply, hopefully that answered your questions xD.
@@mattslams-windows7918 lol my reply was removed
.•*
RAM actually stands for “Read And write Memory” … 😉
It doesn’t - it’s random access memory
stupid 😂
nice ⛑⛑helpful
Thanks!
In the future maybe we will have only one thing.. Cgpu
central graphics processing unit?
Cpu/gpu combo units exist with ARM CPUs.
@@thebtm also with Intel & AMD's APUs