patternMinor
CPU and GPU differences
Viewed 0 times
andcpudifferencesgpu
Problem
What is the difference between a single processing unit of CPU and single processing unit of GPU?
Most places I've come along on the internet cover the high level differences between the two. I want to know what instructions can each perform and how fast are they and how are these processing units integrated in the compete architecture?
It seems like a question with a long answer. So lots of links are fine.
In the CPU, the FPU runs real number operations. How fast are the same operations being done in each GPU core? If fast then why is it fast?
I know my question is very generic but my goal is to have such questions answered.
Most places I've come along on the internet cover the high level differences between the two. I want to know what instructions can each perform and how fast are they and how are these processing units integrated in the compete architecture?
It seems like a question with a long answer. So lots of links are fine.
In the CPU, the FPU runs real number operations. How fast are the same operations being done in each GPU core? If fast then why is it fast?
I know my question is very generic but my goal is to have such questions answered.
Solution
This are not Real numbers as $\mathbb{R}$, but at this point - CPU has double precision floating point numbers, GPU very low number of units processing them, floats on GPU are halfs.
This is due to graphics (this was the main goal before parallel processing), where results are rounded to display, so speed vs accuracy tradeof went that way.
GPU core frequencies are smaller than CPUs, number of operations is very limited on GPU (boosted by video decoder), and there is a huge difference in branch prediction - CPU has very long and complex prediction, while GPU just recently got it added.
Single core on GPU: it is Streaming Multiprocessor (there are about 4 - 16 per card), it includes cuda cores (which is about 32-64), and they work in lock-step, so it differs from CPU threads (not locked).
It is hard to compare like this, but in short - single core on GPU is still parallel unit working slower than CPU core, less memory, registers and instructions than CPU, with very short branching prediction and preferable half floats, nowadays normal floats but having about one-two processing units for double precision, some time ago integer operations were slower on GPU (not onlu by frequency) - but this changed recently.
The same operation on floats - they are slower on GPU than CPU due to frequency.
You might be interested in AMD architecture, Nvidia architecture and Intel architecture to compare instructions set and hardware differences further.
This is due to graphics (this was the main goal before parallel processing), where results are rounded to display, so speed vs accuracy tradeof went that way.
GPU core frequencies are smaller than CPUs, number of operations is very limited on GPU (boosted by video decoder), and there is a huge difference in branch prediction - CPU has very long and complex prediction, while GPU just recently got it added.
Single core on GPU: it is Streaming Multiprocessor (there are about 4 - 16 per card), it includes cuda cores (which is about 32-64), and they work in lock-step, so it differs from CPU threads (not locked).
It is hard to compare like this, but in short - single core on GPU is still parallel unit working slower than CPU core, less memory, registers and instructions than CPU, with very short branching prediction and preferable half floats, nowadays normal floats but having about one-two processing units for double precision, some time ago integer operations were slower on GPU (not onlu by frequency) - but this changed recently.
The same operation on floats - they are slower on GPU than CPU due to frequency.
You might be interested in AMD architecture, Nvidia architecture and Intel architecture to compare instructions set and hardware differences further.
Context
StackExchange Computer Science Q#56082, answer score: 2
Revisions (0)
No revisions yet.