As there are plenty of 4GHz CPUs out there, why can not these be implemented into a GPU?

 

Or why cannot GPUs attain such speeds?

 

allen

As there are plenty of 4GHz CPUs out there, why can not these be implemented into a GPU?

 

Or why cannot GPUs attain such speeds?

 

allen

 

Different architectures. I'm far from knowledgeable on the matter but the x86 architecture is typically 2 - 8 cores while current desktop GPU archs sport thousands of processing units. The pipeline depth is probably much shorter too in GPUs.

GPUs already get quite hot with current clock speeds so there must be a thermal constraint there too

GPUs have a lot more computing power than x86 CPUs, the x86 architecture is pretty ancient but remains current because of backwards compatibility 

GPUs have a lot more computing power than x86 CPUs, the x86 architecture is pretty ancient but remains current because of backwards compatibility 

 

The stock speed 4770K Haswell chip does 177 GFLOPs (random internet search), while a current gen GTX 780 Ti does around 5 TFLOPs.

The stock speed 4770K Haswell chip does 177 GFLOPs (random internet search), while a current gen GTX 780 Ti does around 5 TFLOPs.

 

Yeah, 28 times faster than a 4700K in floating point computing. not too bad

Then the reverse question would be why cannot CPUs catch up with GPUs?

Then the reverse question would be why cannot CPUs catch up with GPUs?

 

x86 CPUs probably can't catch up for the same reasons above. It's an architecture that's like 40 years old so expect it to keep lagging behind GPUs 

Look at how much faster Haswell is compared to Sandy Bridge, 15% give or take. Now compare Fermi vs Kepler: a GTX 780 is like twice as fast as a GTX480. That's a whopping 100%

Emerging architectures like ARMv8 might have a better shot

Desktop CPUs are designed to handle many different types of tasks, such as integer, floating point, SSE etc., while GPUs are highly specialized for graphics work (and other tasks with similar processing needs). Much of the graphics processing work lends itself very well to being highly parallelized, whereas many tasks on a CPU don't. For GPUs this means lots of simpler processing units running in parallel at lower speeds. Trying to run an OS like Windows on a GPU alone would not be very pleasant. And don't get hung up on the x86 ISA being decades old. The internal designs of the CPUs themselves have changed quite significantly from the first 8086 models. Currently the trend by both AMD and Intel, in fact, is to bring the GPU into the CPU itself with the long term goal of using it to replace the floating point module CPUs currently have. So far the x86 series has killed more than a few competing CPUs over the years, including Intel's own Itanium line.

 

To really make things interesting, Intel took several older x86 processing cores and combined them to produce a GPU (which if I recall correctly, is where Phil Taylor went after leaving ACES). It didn't quite work out as a GPU, but has since been developed into a high performance coprocessor under the Xeon Phi brand.

 

https://en.wikipedia.org/wiki/Larrabee_(microarchitecture)

 

https://en.wikipedia.org/wiki/Xeon_Phi

 

At the moment ARM is not comparable at all to x86 CPUs in terms of processing performance as ARM is aimed at low power uses, such as cell phones, and sacrifices processing performance for low power and improved battery life. On the other hand, much of Intel's focus over the past few years has also been on power usage, and they now have low end x86 CPUs that are beginning to compete with ARM's models.