APUs and the future of processing

[Kuzma]

Contents:
Feel free to use the find function (Ctrl+F) in your browser in order to skip to the sections that are most relevant to you.

1.Introduction

1a.APU

1b.AMD APU

1c.Intel APU

1d.Nvidia APU

2a.Integer-based processing

2b.Floating point-based processing

3.The future is APU

 

1.Introduction:
As per request from Vitalius I decided to go out and do some research on the topic of APUs, their current usage, their future usage and how greatly they will impact us in the future. I've compiled everything I've learned into the following thread so sit back and prepare to read, I'm going to try to keep the thread as read friendly as possible so that no matter how much you know about computers you can always take something away from this thread but this topic is quite complex so I don't think I will be able to do that very well

 

(Pictures of the APU die's are all linked on their respective pictures so feel free to click them to compare them and see the similarities and differences)

1a.APU
What is an APU? An APU is an advanced/accelerated processing unit, it's a processing unit with additional processing units alongside the CPU (Central Processing Unit) specialised in other tasks, this can be made up of a CPU and GPU (Graphical Processing Unit) or a different unit specialised for specific processing.

1b.AMD APU

The most commonly known APUs are the APUs from AMD, these are largely advertised as being part of the future of processing and consist of an AMD CPU with an AMD Radeon series GPU; they are a Heterogeneous System Architecture (HSA) and their primary method of using the integrated GPU is the Open Computing Library (OpenCL) and as such require specialised code targetted at GPUs leaving them fairly limited but as it stands they are the most powerful.

1c.Intel APU

One of the less advertised forms of APU are Intel CPUs with Intel HD graphics, although their CPUs are very powerful on their own Intel's CPUs benfit from the Intel HD graphics integrated within them in a very different way than AMD or Nvidia's methods of GPGPU (General processing on Graphics Processing Units) , an example of this is the most recent Intel Quicksync, the usage of Intel HD graphics to process videos and render them faster than the CPU ever could is good accelerated processing; although Intel HD graphics are not very compatible with OpenCL and not compatible at all with CUDA the Intel HD graphics accelerate through the usage of a combination of many small new instruction sets such as AVX.

1d.Nvidia APU

Nvidia's Project Denver, probably the very least known version of an APU is the brain child of a company known as Stexar as an x86-compatible ARM processor, the company was bought out by Nvidia and shortly after they planned to integrate Nvidia GPUs into Project Denver in order to create a new form of APU that would be low power consuption compared to it's x86-based counter-parts with high performance due to CUDA accelerated processing. The project was announced at CES 2011 but x86 compatiblity was stripped due to Nvidia not being granted a license to Intel's patents.

2a.Integer-based processing
Integer-based processing is a type of processing in which programs are restricted to the usage of integers, meaning that there is no use of decimals restricting the number of numbers that can actual be used in the processing but is more suitable for certain situations such as transmitting data in bits as bits can not be divided and as such decimals can not exist. This is the type of processing that CPUs are best at due to their singual structure and not being very friendly to parralelism, they are measured in MIPS (Million Instructions Per Second)

2b.Floating Point-based processing
Floating Point-based processing is the type of processing that we are most used to involving decimals, it is far to complicated to explain here but in short it's the usage of approximations in a precise way in order to increase the speed of processing and ignore seemingly "safe to ignore" data. This is the type of processing that GPUs are best at due to GPUs containing thousands of miniature cores due to the nature of FPOs (Floating Point Operations) being friendly to parralelism, FPOs are measured in FLOPS (FLoating-point Operations Per Second)

3.The future is APU
As mentioned above CPUs specialise in IOPs (Integer Operations) due to the ALUs (Arithematic Logic Unit) within them being much more powerful than the FPUs (Floating Point Units) and as also mentioned above GPUs specialise in FPOs due to their architecture being more friendly with parralelism, the solution seems quite obvious to integrate a GPU within a CPU right? The problem is that not only are there hardware constraints due to the fact that there is limited space within the die, there is also the problem that not every CPU contains a GPU resulting in software having to have specialised code for those who do and those who do not, this is solved in a mixture of ways, OpenCL, CUDA, new instruction sets and Microsoft's DirectCompute the problem is that none of these methods are perfect as it stands but they are indeed all powerful, with OpenCL being the most powerful but having the messiest code, CUDA being the most convenient with very clean and tidy C-style code and DirectCompute having tidy C#-style code but also being the most limited.

As it stands AMD CPUs are leading the way in terms of IOPs as can be seen here , while Intel CPUs are leading in terms of FLOPs as can be seen here however, compared to GPUs which have their FPOs measured in Gigaflops and even then have measurements in the thousands, they are very weak; with AMD being the leaders in terms of ALUs while also having very powerful FPUs in their APUs it seems that they are prepared for the future, the future is currently looking like HSAs will dominate with powerful ALUs and equally powerful FPUs in th form of GPUs to do the computing; the only current bottleneck for them as it stands is software and once that is harmonised HSAs will have no real competition.

Conclusion:

I hope this thread has helped you in some way shape or form, it's very generalised so I can't really say anything specific but next time you buy a CPU (or possibly even an APU ) think about if you're ready for the future...

If you feel I've missed anything or need any help understanding something specfically feel free to PM me and I'll be sure to answer you as soon as possible.

[megadarkwood]

Words, words, WORDS!!!

TL;DR

I skimmed through it, and APUs are amazing! (I have one in my laptop!)

[Bluejay0]

Very detailed article and very well thought. Loved the organization format and research into every section. Excellent work!

[TechFan@ic]

Very informative thank you very much.

[Dravic]

Contents:

Feel free to use the find function (Ctrl+F) in your browser in order to skip to the sections that are most relevant to you.

1.Introduction

1a.APU

1b.AMD APU

1c.Intel APU

1d.Nvidia APU

2a.Integer-based processing

2b.Floating point-based processing

3.The future is APU

 

1.Introduction:

As per request from Vitalius I decided to go out and do some research on the topic of APUs, their current usage, their future usage and how greatly they will impact us in the future. I've compiled everything I've learned into the following thread so sit back and prepare to read, I'm going to try to keep the thread as read friendly as possible so that no matter how much you know about computers you can always take something away from this thread but this topic is quite complex so I don't think I will be able to do that very well

 

(Pictures of the APU die's are all linked on their respective pictures so feel free to click them to compare them and see the similarities and differences)

1a.APU

What is an APU? An APU is an advanced/accelerated processing unit, it's a processing unit with additional processing units alongside the CPU (Central Processing Unit) specialised in other tasks, this can be made up of a CPU and GPU (Graphical Processing Unit) or a different unit specialised for specific processing.

1b.AMD APU

The most commonly known APUs are the APUs from AMD, these are largely advertised as being part of the future of processing and consist of an AMD CPU with an AMD Radeon series GPU; they are a Heterogeneous System Architecture (HSA) and their primary method of using the integrated GPU is the Open Computing Library (OpenCL) and as such require specialised code targetted at GPUs leaving them fairly limited but as it stands they are the most powerful.

1c.Intel APU

One of the less advertised forms of APU are Intel CPUs with Intel HD graphics, although their CPUs are very powerful on their own Intel's CPUs benfit from the Intel HD graphics integrated within them in a very different way than AMD or Nvidia's methods of GPGPU (General processing on Graphics Processing Units) , an example of this is the most recent Intel Quicksync, the usage of Intel HD graphics to process videos and render them faster than the CPU ever could is good accelerated processing; although Intel HD graphics are not very compatible with OpenCL and not compatible at all with CUDA the Intel HD graphics accelerate through the usage of a combination of many small new instruction sets such as AVX.

1d.Nvidia APU

Nvidia's Project Denver, probably the very least known version of an APU is the brain child of a company known as Stexar as an x86-compatible ARM processor, the company was bought out by Nvidia and shortly after they planned to integrate Nvidia GPUs into Project Denver in order to create a new form of APU that would be low power consuption compared to it's x86-based counter-parts with high performance due to CUDA accelerated processing. The project was announced at CES 2011 but x86 compatiblity was stripped due to Nvidia not being granted a license to Intel's patents.

2a.Integer-based processing

Integer-based processing is a type of processing in which programs are restricted to the usage of integers, meaning that there is no use of decimals restricting the number of numbers that can actual be used in the processing but is more suitable for certain situations such as transmitting data in bits as bits can not be divided and as such decimals can not exist. This is the type of processing that CPUs are best at due to their singual structure and not being very friendly to parralelism, they are measured in MIPS (Million Instructions Per Second)

2b.Floating Point-based processing

Floating Point-based processing is the type of processing that we are most used to involving decimals, it is far to complicated to explain here but in short it's the usage of approximations in a precise way in order to increase the speed of processing and ignore seemingly "safe to ignore" data. This is the type of processing that GPUs are best at due to GPUs containing thousands of miniature cores due to the nature of FPOs (Floating Point Operations) being friendly to parralelism, FPOs are measured in FLOPS (FLoating-point Operations Per Second)

3.The future is APU

As mentioned above CPUs specialise in IOPs (Integer Operations) due to the ALUs (Arithematic Logic Unit) within them being much more powerful than the FPUs (Floating Point Units) and as also mentioned above GPUs specialise in FPOs due to their architecture being more friendly with parralelism, the solution seems quite obvious to integrate a GPU within a CPU right? The problem is that not only are there hardware constraints due to the fact that there is limited space within the die, there is also the problem that not every CPU contains a GPU resulting in software having to have specialised code for those who do and those who do not, this is solved in a mixture of ways, OpenCL, CUDA, new instruction sets and Microsoft's DirectCompute the problem is that none of these methods are perfect as it stands but they are indeed all powerful, with OpenCL being the most powerful but having the messiest code, CUDA being the most convenient with very clean and tidy C-style code and DirectCompute having tidy C#-style code but also being the most limited.

As it stands AMD CPUs are leading the way in terms of IOPs as can be seen here , while Intel CPUs are leading in terms of FLOPs as can be seen here however, compared to GPUs which have their FPOs measured in Gigaflops and even then have measurements in the thousands, they are very weak; with AMD being the leaders in terms of ALUs while also having very powerful FPUs in their APUs it seems that they are prepared for the future, the future is currently looking like HSAs will dominate with powerful ALUs and equally powerful FPUs in th form of GPUs to do the computing; the only current bottleneck for them as it stands is software and once that is harmonised HSAs will have no real competition.

Conclusion:

I hope this thread has helped you in some way shape or form, it's very generalised so I can't really say anything specific but next time you buy a CPU (or possibly even an APU ) think about if you're ready for the future...

If you feel I've missed anything or need any help understanding something specfically feel free to PM me and I'll be sure to answer you as soon as possible.

 

Let's clarify something.

 

Current APUs are shit. They have nothing to do with future you are describing and if (when) the future you are describing comes, those current APUs will still be... shit. Those aren't APUs you are looking for folks. AMD doesn't have them... yet.

 

So buying an APU right now is the worst kind of waste you can do, they have horrible AND I MEAN HORRIBLE price-to-performance, few more bucks and you got a MUCH more powerful CPU+GPU which, what's even better, you can UPGRADE (that's right, try to upgrade APU on FM2 mobo which is dead socket, and even if you could, what you gonna do? upgrade from APU to APU? LOL that's double wasted money, at this point you'd afford a better CPU and much more powerful GPU if you paid that money initially and got much better performance).

[Benjio]

Let's clarify something.

 

Current APUs are shit. They have nothing to do with future you are describing and if (when) the future you are describing comes, those current APUs will still be... shit. Those aren't APUs you are looking for folks. AMD doesn't have them... yet.

 

So buying an APU right now is the worst kind of waste you can do, they have horrible AND I MEAN HORRIBLE price-to-performance, few more bucks and you got a MUCH more powerful CPU+GPU which, what's even better, you can UPGRADE (that's right, try to upgrade APU on FM2 mobo which is dead socket, and even if you could, what you gonna do? upgrade from APU to APU? LOL that's double wasted money, at this point you'd afford a better CPU and much more powerful GPU if you paid that money initially and got much better performance).

 

I have an A10-5800k and I couldn't be happier with it, plays all the games I play at pretty much max settings for only $120, so don't be so hateful of them please...

[Kuzma]

Very informative thank you very much.

Thank you just to say, your thread title of "How Far Ahead Intel Really Is?" is untrue since they're kinda behind if you take APUs into account but

[Dravic]

I have an A10-5800k and I couldn't be happier with it, plays all the games I play at pretty much max settings for only $120, so don't be so hateful of them please...

 

 

[Benjio]

 

 

That's not really the point... I get that if you pay a little more you can get much better performance, an APU is all I need though. For people who don't game much and don't need an external GPU, APUs are perfect, especially for things like HTPCs and simple web-browsing computers.

[Dravic]

That's not really the point... I get that if you pay a little more you can get much better performance, an APU is all I need though. For people who don't game much and don't need an external GPU, APUs are perfect, especially for things like HTPCs and simple web-browsing computers.

They are perfect if they want the bottom of the performance for price of a good set.

 

Tell me how much is a good CPU? well if you get your hands on phenom II x4 965 for around 80$... this is about 70$ GPU (1GB GDDR5 of course). 150$ and kills APU with memory bandwidth and performance. The mobo is around same price for FM2 and AM3+ but with AM3 you can upgrade to FX-8350 if thats your thing. Or you can get an old good Pentium (dual core) sandy/ivy and get CHEAP ASS 1155 mobo and that's even less (but you can upgrade to i5 in the future!).

 

And you can upgrade GPU if you need without changing whole platform/CPU as well.

[Vitalius]

That's not really the point... I get that if you pay a little more you can get much better performance, an APU is all I need though. For people who don't game much and don't need an external GPU, APUs are perfect, especially for things like HTPCs and simple web-browsing computers.

This.

@Kuzma,

 

Great post.

SPOILERS

Also, you have to use @ or I do not see it. Thanks though, great read. I enjoyed it. Glad I could give you a topic for another awesome and informative thread.

[Kuzma]

They are perfect if they want the bottom of the performance for price of a good set.

 

Tell me how much is a good CPU? well if you get your hands on phenom II x4 965 for around 80$... this is about 70$ GPU (1GB GDDR5 of course). 150$ and kills APU with memory bandwidth and performance. The mobo is around same price for FM2 and AM3+ but with AM3 you can upgrade to FX-8350 if thats your thing. Or you can get an old good Pentium (dual core) sandy/ivy and get CHEAP ASS 1155 mobo and that's even less (but you can upgrade to i5 in the future!).

 

And you can upgrade GPU if you need without changing whole platform/CPU as well.

A dual core pentium will get ripped to shreads in anything multi-threaded by an a10-6800k & to be honest prices are so variable all over the world you can't really say "you could have got x" anymore >_<

[Benjio]

They are perfect if they want the bottom of the performance for price of a good set.

 

Tell me how much is a good CPU? well if you get your hands on phenom II x4 965 for around 80$... this is about 70$ GPU (1GB GDDR5 of course). 150$ and kills APU with memory bandwidth and performance. The mobo is around same price for FM2 and AM3+ but with AM3 you can upgrade to FX-8350 if thats your thing. Or you can get an old good Pentium (dual core) sandy/ivy and get CHEAP ASS 1155 mobo and that's even less (but you can upgrade to i5 in the future!).

 

And you can upgrade GPU if you need without changing whole platform/CPU as well.

 

Its not like I can't add an external GPU in if I want. An APU isn't too much more than a CPU with similar performance, and its always nice to have the option to use onboard if you're graphics card dies or something of the sort.

[Vitalius]

Its not like I can't add an external GPU in if I want. An APU isn't too much more than a CPU with similar performance, and its always nice to have the option to use onboard if you're graphics card dies or something of the sort.

That, and Hybrid Crossfire (Crossfire between an APU and a dedicated GPU such as a 7770) is a thing. Just saying. Sure it does not give much upgrade ability, but it gives you some which is good enough for the type of people who would buy an APU (those who only game a bit).

[Dravic]

That, and Hybrid Crossfire (Crossfire between an APU and a dedicated GPU such as a 7770) is a thing. Just saying. Sure it does not give much upgrade ability, but it gives you some which is good enough for the type of people who would buy an APU (those who only game a bit).

 

Yes, it's a thing that gives you 10-20% performance boost for a price of full GPU. everything else is just runt/drop frames and you are not seeing what fraps says you should see. Tom's hardware tested it.

 

Besides APU is nothing else than FX-4300. You kidding me? Who in this world would buy FX-4300 with just 2 FPUs (but yes, oh yes it has 4 alus it must be goooooooood...) and try to run a GPU off it? It'd have shitty performance in games anyway. At least Phenom II x4 965. It's still a bit better overally than FX-4300 thanks to full cores.

[Kuzma]

Yes, it's a thing that gives you 10-20% performance boost for a price of full GPU. everything else is just runt/drop frames and you are not seeing what fraps says you should see. Tom's hardware tested it.

 

Besides APU is nothing else than FX-4300. You kidding me? Who in this world would buy FX-4300 with just 2 FPUs (but yes, oh yes it has 4 alus it must be goooooooood...) and try to run a GPU off it? It'd have shitty performance in games anyway. At least Phenom II x4 965. It's still a bit better overally than FX-4300 thanks to full cores.

Considering an a10-6800k out performs everything else AMD have single core I think that speaks for itself, I think the fact that if you're playing a game and want to render out a yt video, you can offload the video onto your internal gpu is quite a nice thing; the fact you can play games with no performance hit.

[slasher0161]

Apu's for anyone who doesn't play cpu intensive games and have a limited budget can be quite good (especially if they are intending to pick up a dedicated card later on, it does happen).

 

When assembling my htpc I ran a few tests with my A8-6600k and chucked my 7870xt with it just to see what it could do. Playing all games at max detail the only game that became unplayable was skyrim however that was with heavy modding, even WoW was playable outside of 25m content (however 25m content makes even 4.5ghz i5's take a hit).

 

Here is a screen shot of the sleeping dogs benchmark to give you an idea.

 

[TechFan@ic]

Thank you just to say, your thread title of "How Far Ahead Intel Really Is?" is untrue since they're kinda behind if you take APUs into account but

Wanted to name my thread after a wide-spread misconception to spark interest.

[Kuzma]

Wanted to name my thread after a wide-spread misconception to spark interest.

touché

[leipero]

I would add something also, all well explained, but, most post i did read here miss the point entirely, and Kuzma explained it well. With APU, you can use CPU cores for processing one sort of code, while parallel you can use GPU for processing another type of code, and on top of that, use GPU for actuall Graphic (if we speak about gaming), That way, you get incredibly fast "CPU" (in fact APU) with fast GPU win-win situation.

 

Few problems with that, in my opinion. First is the software, but no one will bother with software if there is no eco-system to run at..., that's why AMD push it's own APU's. 2dn thing is one and only L3 cache, it may seem irelevant, but even on piledriver/storm... architecture, it plays crucial role, even tho i don't know even close, let alone more than AMD engineers, i bet they think the same, only problem is, it is not cost-effective, while it is beneficial enough to implement.

 

If AMD follows "less BUS traffic = better" logic, have no doubts that in future APU's L3 cache will be present, especially when you ahve on mind that same bus now have to transfer data from both GPU and CPU (inside of APU) on "one clock". That might be (or it is) solved with separate roots, but it is way too complicated and not effective. So in short, more work on APU, less transfers between APU and memory = better. Problem is, GPU on APU's don't have their own memory, they use system RAM, and that is only bad thing about APU's. Implementing large cache for GPU inside of APU is maybe out of the question, but for small "CPU alike" FPU operations, small dedicated cache would make much more sense. It is question if it would be beneficial enough to be cost effective, but basically, I'm completely out of topic, and i was "thinking out loud", so ignore this last paragraph .

[Kuzma]

I would add something also, all well explained, but, most post i did read here miss the point entirely, and Kuzma explained it well. With APU, you can use CPU cores for processing one sort of code, while parallel you can use GPU for processing another type of code, and on top of that, use GPU for actuall Graphic (if we speak about gaming), That way, you get incredibly fast "CPU" (in fact APU) with fast GPU win-win situation.

 

Few problems with that, in my opinion. First is the software, but no one will bother with software if there is no eco-system to run at..., that's why AMD push it's own APU's. 2dn thing is one and only L3 cache, it may seem irelevant, but even on piledriver/storm... architecture, it plays crucial role, even tho i don't know even close, let alone more than AMD engineers, i bet they think the same, only problem is, it is not cost-effective, while it is beneficial enough to implement.

 

If AMD follows "less BUS traffic = better" logic, have no doubts that in future APU's L3 cache will be present, especially when you ahve on mind that same bus now have to transfer data from both GPU and CPU (inside of APU) on "one clock". That might be (or it is) solved with separate roots, but it is way too complicated and not effective. So in short, more work on APU, less transfers between APU and memory = better. Problem is, GPU on APU's don't have their own memory, they use system RAM, and that is only bad thing about APU's. Implementing large cache for GPU inside of APU is maybe out of the question, but for small "CPU alike" FPU operations, small dedicated cache would make much more sense. It is question if it would be beneficial enough to be cost effective, but basically, I'm completely out of topic, and i was "thinking out loud", so ignore this last paragraph .

maybe 32MB of L3 cache for the GPU would be an amazing idea for lower latency and better performance but I guess with HUMA and the gpu being able to access the RAM without having to go through the CPU first is amazing and will give probably 10-15% performance increase per 300mhz

[leipero]

maybe 32MB of L3 cache for the GPU would be an amazing idea for lower latency and better performance but I guess with HUMA and the gpu being able to access the RAM without having to go through the CPU first is amazing and will give probably 10-15% performance increase per 300mhz

Yeah, maybe that is good enough solution, but again, DMA conflicts etc... , if AMD goes that way, they will need to apply strict quality control over board partners, even if that means higher cost for motherboards. For example, on my motherboard, i tend toa void DMA as much as possible (used only for GPU) because of all nasty effects when conflict arrives, if AMD is aware of that, and can solve it, why not using RAM, I;m not against as long as quality is in place .

[Kuzma]

Yeah, maybe that is good enough solution, but again, DMA conflicts etc... , if AMD goes that way, they will need to apply strict quality control over board partners, even if that means higher cost for motherboards. For example, on my motherboard, i tend toa void DMA as much as possible (used only for GPU) because of all nasty effects when conflict arrives, if AMD is aware of that, and can solve it, why not using RAM, I;m not against as long as quality is in place .

Fair enough, well to be honest I think APUs will benfit from DDR4 the most out of anything, which is why Intel have suddenly upped the power of their own Integrated graphics in prepartion for Skylake and DDR4; I think Intel knows that at the current point in time they are actually largly behind AMD, even though they are dominating the current markets, it's going to be a second coming of the Athlon - I don't think they'll get away with paying companies to use primarily their CPUs anymore.

[leipero]

Fair enough, well to be honest I think APUs will benfit from DDR4 the most out of anything, which is why Intel have suddenly upped the power of their own Integrated graphics in prepartion for Skylake and DDR4; I think Intel knows that at the current point in time they are actually largly behind AMD, even though they are dominating the current markets, it's going to be a second coming of the Athlon - I don't think they'll get away with paying companies to use primarily their CPUs anymore.

I actually just did read article about XBOX One, and it seems they already done that, "APU" inside of XBOX One have 47MB of memory. Here is the quote:

 

"The chip covers an area of 363 square millimeters and is built in a 28-nanometer state-of-the-art manufacturing process. It has 5 billion basic components, known as transistors, on a single chip that combines a microprocessor, graphics processor, and 47MB of memory."

 

Source Guru3D.

[Kuzma]

I actually just did read article about XBOX One, and it seems they already done that, "APU" inside of XBOX One have 47MB of memory. Here is the quote:

 

"The chip covers an area of 363 square millimeters and is built in a 28-nanometer state-of-the-art manufacturing process. It has 5 billion basic components, known as transistors, on a single chip that combines a microprocessor, graphics processor, and 47MB of memory."

 

Source Guru3D.

I did indeed read that but I kinda assumed it was DDR3 memory it never occurred to me it might be cache.