APU vs (GPU+CPU)

[wajed]

Only performance wise:-

What is the difference between a CPU (say first generation core i7) combined with a discrete GPU (say HD6570) AND an APU that consists of the same core i7 CPU and the same HD6570 GPU?

Please, if you can, elaborate rather than giving a "this is better" answer.

 

[AtenRa]

The only difference will be in memory bandwidth with the Discrete GPU(CPU + GPU) to be the winner (Higher Performance)

 

[Arkadrel]

What is a APU:
http://www.youtube.com/user/AMDUnprocessed#p/a/u/0/BihrG7DhhBM


They remove a bottle neck, of the interconnect bus between the CPU and GPU... so besically theres a big gain when it comes to software that uses GPGPU (this is just from listning to the video). The idea is to kill CPU only programs, and make everything benefit from haveing a GPU as well (where its reasonable compaired to gains in performance), which will result in massive improvements in compute power for more or less everyone.

Another benefit seems to be the costs, they can radically lower the cost, by moveing things onto 1 chip and do so while useing Lower power too.

http://sites.amd.com/PublishingImag...PNG/AMD_E-350_Accelerated_Processor_Chart.png


So instead of ending up with a CPU for 80$ + grafic card for 80$, they can sell a CPU+GPU 2in1 chip for say 85$ that can do the same as above(cpu+ discrete gpu).
That way the end user, the consumer can save like 75$ by going this route, without it costing him performance compaired to cpu+discrete gpu.

Ofc these cards wont be competeing at enthusiast levels..... they ll probably stay low end, for along time. So cutting egde, fastest fps ect in games, wont be from a APU. You ll still need to buy a 400-600$ discrete grafics card for that.

However once llano comes out, anything slower than a 440-450/5670 wont make sense to buy, if your makeing a new pc, because it ll come along on the APU.

 

[wajed]

I understand the cut in cost (well, I didn't imagine it would be that much,) and I understand it's more power efficient.

If it's all about GPGPU, I would wait for Trinity, because Liano won't be beneficial for GPGPU, and for now it's not the best option or the second best option for CPU tasks.

"The only difference will be in memory bandwidth with the Discrete GPU(CPU + GPU) to be the winner (Higher Performance)"
Are there any figures showing how much the difference is?


EDIT: "because Liano won't be beneficial for GPGPU," < I mean there aren't much GPGPU everyday apps now, so when it comes to GPGPU Liano won't be beneficial.

 

[Idontcare]

Only performance wise:-

What is the difference between a CPU (say first generation core i7) combined with a discrete GPU (say HD6570) AND an APU that consists of the same core i7 CPU and the same HD6570 GPU?

Please, if you can, elaborate rather than giving a "this is better" answer.

Your question places an unnecessary limitation/restriction on what your APU is going to be.

An APU could be nothing more than a GPU bolted onto the same piece of silicon that the CPU is bolted onto...but that isn't what Fusion or an APU is all about.

The idea with making an APU is that you meld the ISA's of both architectures in some fashion - the extent of which is up to the design goals and the budget - like adding SSE4.1 to an existing x86 ISA and architecture.

The benefits of an APU are not to enhance the existing features or functions of the existing components in the existing standalone products. Just as SSE4.1 does not make the existing x87 ISA perform better, but rather it provides an alternative that is faster and better, making the prior methods all but obsolete excepting for maintaining backwards compatibility with legacy apps.

If you are looking at APU's as "they done bolted my GPU into my CPU right here, now it goes to teh 11!" then you will be disappointed.

If you are looking at APU's as "can't wait for the next generation of software to get here which will have been optimized and compiled to take advantage of the next-gen hardware feature set! w00t!" then you have appropriately aligned your expectations with reality.

Look at this as being similar to the benefits of adding the FPU to the x86 chip back in the 486 days, or adding MMX or SSE extensions to existing x86 ISA. This isn't like adding more cache or an integrated memory controller which benefited all existing instructions, this is new stuff that will supplant existing stuff.

 

[IntelUser2000]

Before the focus was overly towards one metric. The CPU guys go, "CPUs are the best, and its the only way to go", and the GPU guys say "GPUs can replace everything and anything".

Same with more cores vs single thread. One guy says we need hundreds and thousands of cores, and the other says all we need is single thread.

Now there's a new kid on the block saying "We should combine the two for to use the best of both".

This is like tech crowd way of saying "Why can't we stop fighting and get along?"

Really, everything has advantages and disadvantages. Looking at ways to find better performance should never be limited to one metric, which is something opposite of what marketing guys wants us to think.

However once llano comes out, anything slower than a 430/5570 wont make sense to buy, if your makeing a new pc, because it ll come along on the APU.

Corrected the above.

 

[Arkadrel]

I understand the cut in cost (well, I didn't imagine it would be that much,) and I understand it's more power efficient.

If it's all about GPGPU, I would wait for Trinity, because Liano won't be beneficial for GPGPU, and for now it's not the best option or the second best option for CPU tasks.

"The only difference will be in memory bandwidth with the Discrete GPU(CPU + GPU) to be the winner (Higher Performance)"
Are there any figures showing how much the difference is?

EDIT: "because Liano won't be beneficial for GPGPU," < I mean there aren't much GPGPU everyday apps now, so when it comes to GPGPU Liano won't be beneficial.

Current gen of APUs will likely be close to memory bound....

Lets say you use dual channel, DDR3 1866 ram.... the memory bandwidth limit would be around:

DDR3-1866 = 64x1.866 / 8 x2 = 29,856 GB/s (in dual channel mode) (I think this is how the math is)

A Radeon HD 5670 with GDDR3 has 25.6 GB/s Memory bandwidth.
A Radeon HD 6970 with GDDR5 has 176 GB/s Memory bandwidth.

Now Memory bandwidth ISNT the only thing that dictates how fast a grafics card is...
So we ll probably see the LLANO performing around the ~5670 level, maybe a little above or a little below it.

Most Grafics cards have higher memory bandwidth than this (28 GB/s), so obviously useing the main systems memory is only gonna hold for this gen of APUs (unless improvements are made to mem bandwidth for the system, say DDR4 or tripple/quad channel ram).



Solution to memory bandwidth issues:
Next gen APU will probably have a "extra" ram slot, where you can put in some GDDR5 next to the cpu or motherboards will come with GDDR5 Sideport memory. OR by then main system memory will all be DDR4 with tripple channel or quad channel or something to give more system memory bandwidth.

OR they add some eDRAM to the proccessor to give it memory bandwidth for the GPU.
IBM's POWER7 processor, Sony's PlayStation 2, Sony's PlayStation Portable, Nintendo's GameCube, Nintendo's Wii, Apple Inc.'s iPhone, Microsoft's Zune HD, and Microsoft's Xbox 360. ALL use this... time the mainstream pcs do it too?

The 2nd gen of APUs is gonna be what makes or breaks it... if they figour out a way around the Memory bandwidth issues,.. we ll likely see the power of the GPUs inside the APUs skyrocket.

 

[gregoryvg]

The bandwidth may be there on the build in GPU's, but the horsepower isn't. A 5670 would blow away any of the on-die GPU's.

 

[Arkadrel]

The bandwidth may be there on the build in GPU's, but the horsepower isn't. A 5670 would blow away any of the on-die GPU's.

a 5670 has 400 shader units, running 775 mhz.
a llano gpu, is rumored to have 480 shader units, no idea of core clock.



Ive heard alot say between 5500-5600 (from sites that have any news on the llano), and Im guessing that ll depend on what ram you couple it up with.
HD5570-HD5670.... Im guessing closer to the 5670 (if you give it 1866 DDR3 ram dual channel).


You seen what the Zacate can do, with 17-18watts, and a 80 shader units (on 40nm node). A 480 shader unit, with up to 95watts is bound to be alot faster (on 32nm node).


AMD mentioned over 500+ GFlops as given for Llano by AMD itself in one of those youtube videos I believe (doesnt mean its 500, just mean it has atleast 500+ Gflops). A HD5670 has ~620 GFlops.. I think it ll be close, or atleast faster than the 5570.

 

[LiuKangBakinPie]

Its all about the scaling.

Higher the resolutions -----> Less emphasis on the CPU more on the GPU
Lower resolutions --------> More on the cpu less on the gpu.

So lower resolutions the faster I7 will outperform the Amd counterparts. But as soon as you up it the gap closes and the gpu quality takes over

 

[Arkadrel]

@ LioKangBakinPie

Im not so sure.... the HD 3000 from the 2600k, is gonna be like 2-3 times slower (if the Llano GPU is like a 5550, and its probably faster)... thats a huge handicap for sandy bridge to make up for in cpu strength. No doubt the Sandybridge is gonna out perform the Llano in CPU stuff (can intels GPU do GPGPU?) in stuff thats cpu intensive, but same time anything with lots of grafics, the Llano is gonna outperform the Sandybrigdes HD3000 GPU.

 

[Mopetar]

That's to be expected considering that the first Llano parts will be based off of the Stars core architecture and those cores don't perform as well Intel's most recent offerings. The GPU portion of the chip will demolish Intel's offerings however. It will probably become a more capable processor once they start using the bulldozer cores, which while they may not be as good as what Intel has, should be a lot better than what AMD has.

I imagine that Llano will make a good chip for people interested in casual gaming. Based on the components that will be going into it, it should handle 1680 x 1050 well at an affordable cost. It should also improve performance for some applications that can take advantage of the GPU.

Llano will be targeted at the mainstream, people who want to play some games and use their machine for moderate workloads at a reasonable price. Bulldozer will be more geared towards professionals or the gamer market that is going to use discrete graphics anyways.

What confuses me the most is that if you consider what I just said, Intel's position is completely bizarre. With Intel's desktop SB offerings, the K series parts have the best graphics, but there's currently no motherboard that allows you to take advantage of the onboard graphics and overclock the CPU at the same time. More confusing is that the K series parts were given the best graphics when these are the parts most likely to be paired with a discrete card.

 

[Concillian]

I know 2-3 people waiting on Llano.

Just as Mopetar explained, they are casual gamers. I take it a step further, and add that they are casual gamers who prefer aesthetics to performance. They just want something that performs like a 5450 or 5550, but they don't want to spend for it and they want a cute shiny, tiny Dell case or an all-in one style monitor / case.

Llano is perfect for these people, who are also price sensitive enough that anythign SB based is out of current budget, primarily because anything with decent on-die graphics for the desktop (K chip) has rather insane markups over what they need (an i3 CPU would be fine for them.)

Llano may not be the fastest around, but I think there is a good market for it. By keeping the HD3000 only on the K chips, Intel has basically handed the bulk of the budget, casual gamer market who would want an APU to Llano (at least among people who have done any research into the matter). Can't wait to see what the software guys start doing with APUs. It'll be an interesting next few years. The hardware side is always the easy part. Takes lots of time for the software guys to figure out what's the optimal way to take advantage of that hardware.

 

[wajed]

Intel doesn't have the upper hand at GPU performance (Isn't Sandy Bridge's GPU only 1/3 as powerful as Liano's?)
So, I think it is beneficial for Intel that it doesn't implement GPGPU right now, since software developers won't be very excited to redesign there software for small portion of consumers.

That will only give them little time, since Trinity is on the way, and it won't give Intel the chance to stay the dominant if it doesn't have superior CPU and comparable GPU.

 

[AtenRa]

However once llano comes out, anything slower than a 440-450/5670 wont make sense to buy, if your makeing a new pc, because it ll come along on the APU.

Not all of us will buy an APU, Bulldozer will not be an APU and even a HD5450 could be used with it, same goes for SandyBridge 2011 or anyone else who would want to install a new gen low end GPU to there system, low end GPUs will continue to exist long after Llano's release.

 

[Mopetar]

Intel doesn't have the upper hand at GPU performance (Isn't Sandy Bridge's GPU only 1/3 as powerful as Liano's?)
So, I think it is beneficial for Intel that it doesn't implement GPGPU right now, since software developers won't be very excited to redesign there software for small portion of consumers.

That will only give them little time, since Trinity is on the way, and it won't give Intel the chance to stay the dominant if it doesn't have superior CPU and comparable GPU.

Part of the problem is that Intel doesn't have support for DX 11 or OpenCL with their current IGP which holds them back in a few ways. Without support for the latest technologies, it's more of a pain to target their CPUs for GPGPU.

We don't have performance numbers for Bulldozer yet, and we can only speculate about Llano so it's hard to say how Intel will react. We'll assume that they're going to be ahead on the CPU side of things, but they're much farther behind on the GPU side of things.

 

[Mopetar]

I take it a step further, and add that they are casual gamers who prefer aesthetics to performance. They just want something that performs like a 5450 or 5550, but they don't want to spend for it and they want a cute shiny, tiny Dell case or an all-in one style monitor / case.

This is probably where Llano is going to excel. I honestly wouldn't be surprised if Apple ditched Intel and started using Llano in their Mac Mini's. Intel's 2100T provides 2 cores with hyper-threading, and integrated graphics all in a 35W TDP. AMD should be able to provide something similar, but with vastly superior graphics that have better OpenCL support.

 

[Concillian]

Even creating apps that take advantage of Sandy Bridge EUs is more than we have right now. The fatal flaw in SB is that there is no allowance to use the EUs if you have a discrete card, meaning you can't encourage software developers to do something like put physics on the EUs and games on the discrete GPU. You can't even encourage mainstream apps to use the EUs because a subset of users will be using discrete cards and not have access to their own hardware.

I hope AMD is not so shortsighted on the GPU portion of their APUs. All indications are that they are not, but without anything released yet, nothing is for certain.

 

[Mopetar]

Not all of us will buy an APU, Bulldozer will not be an APU and even a HD5450 could be used with it, same goes for SandyBridge 2011 or anyone else who would want to install a new gen low end GPU to there system, low end GPUs will continue to exist long after Llano's release.

The low end of the low end will most likely disappear. Of course there will always be a low end market simply because previous mid-range parts will remain unsold and eventually drop in price.

Most people will just end up using whatever graphics are on the motherboard rather than getting a discrete card if they have absolutely no use for a GPU or IGP.

I don't foresee AMD making a large number of discrete cards that aren't sufficiently better than whatever they're putting in their processors. The only way that would make any sense is if they released some kind of ability to crossfire the IGP in the CPU with a discrete card, which I don't think will happen.

The only other reason they would continue to make low-end graphics cards is because Intel's graphics continue to suck compared to AMD/Nvidia and there's a market for such cards in HTPC setups.

 

[LiuKangBakinPie]

Yes advantage in a benchmark WHOOPEE!!!

 

[Concillian]

We don't have performance numbers for Bulldozer yet, and we can only speculate about Llano so it's hard to say how Intel will react. We'll assume that they're going to be ahead on the CPU side of things, but they're much farther behind on the GPU side of things.

I don't think there's a lot of speculation on Llano. We know it's not going to be offering much, if any IPC improvement. We know the available memory bandwidth will be the limitation on GPU performance, so it has a fairly fixed ceiling. The only real question on performance is clock speeds that will be achievable on their 32nm process. CPU performance is going to be on-par (at a given clock) with current AthlonIIs, maybe slightly better. And GPU speeds will be capped by memory bandwidth to 5550/5570 levels.

The things we don't know is how it will balance power consumption. This is a weak spot of AMD that they have had slides showing that they acknowledge and plan to correct with Llano, it will also be 32nm, so that will be a boost too. I don't expect it to compete directly with any SB on CPU power per watt, but it should be a lot closer than their current offerings. Good enough that it's not a major issue for the desktop anyway.

The other question is the particulars of implementation of the APU. Drivers and hybrid functions with discrete cards are things AMD has a clear potential on, and it would make sense for them to put a lot of focus here. This is how it can make up for the areas where nobody expects them to be able to catch up to Intel this generation. If you can't compete directly, if you create a niche and exploit that niche, you can still maintain decent profitability.

IMO the success of the niche that Llano fits into hinges on drivers and getting software on-board utilizing the GPU portion of the CPU for more than just games ASAP.

Can you imagine, for example, if they got into bed with Adobe, and had Photoshop or Premiere using the GPU portion of the APU for accelerating software like that? That's the kind of thing that would be a home run. I don't expect that to happen, but can you imagine if it did? That would really shake things up, as it would potentially create an a path where some businesses have a real reason to want something other than what they always buy, because that's what they've always done (Intel).

 

[Arkadrel]

Even creating apps that take advantage of Sandy Bridge EUs is more than we have right now. The fatal flaw in SB is that there is no allowance to use the EUs if you have a discrete card, meaning you can't encourage software developers to do something like put physics on the EUs and games on the discrete GPU. You can't even encourage mainstream apps to use the EUs because a subset of users will be using discrete cards and not have access to their own hardware.

I hope AMD is not so shortsighted on the GPU portion of their APUs. All indications are that they are not, but without anything released yet, nothing is for certain.

As I understand it, reguardless of which grafics card (amd) you put into your pc, the GPGPU funktions of the IGP GPU of the LLANO will stack with those of the Discrete GPU.

Makeing LLano + GPU = GPGPU monster computer.
(which is the idea behinde the APUs, to push GPGPU computeing to the next stage)

I don't foresee AMD making a large number of discrete cards that aren't sufficiently better than whatever they're putting in their processors. The only way that would make any sense is if they released some kind of ability to crossfire the IGP in the CPU with a discrete card, which I don't think will happen.

They will, you can Crossfire your Llano with a 5570/6xxx ish type card.
Llano + 5570/6xxx = ~5750 performance once crossfired (my guess)

which will be a cheapish way of getting low res gameing to people.

Can you imagine, for example, if they got into bed with Adobe, and had Photoshop or Premiere using the GPU portion of the APU for accelerating software like that?

Im 99&#37; sure amd knows thats what it needs for APU, with their gpgpu strengths, to sell.
Id say chances of this *not* happending would surprise me. This would mean Adobe would have more that could use Adobe (more to sell it too), so why would they not want it as well? Amd would have a easier time selling APUs. This is bound to happen.

 

[Mopetar]

They will, you can Crossfire your Llano with a 5570/6xxx ish type card.
Llano + 5570/6xxx = ~5750 performance once crossfired (my guess)

which will be a cheapish way of getting low res gameing to people.

Nice. I wasn't aware of that.

 

[drizek]

You could already do Crossfire since 780G and a 3200/3400, IIRC.

Pretty useless, although people with real video cards should be able to take advantage of Llano to get access to the motherboards video connectors, for Eyefinity or if their video board manufacturer was dumb enough not to include Displayport(*ahem* GTX 460).

 

[PreferLinux]

Id say chances of this *not* happending would surprise me. This would mean Adobe would have more that could use Adobe (more to sell it too), so why would they not want it as well? Amd would have a easier time selling APUs. This is bound to happen.

Currently Adobe uses CUDA in Photoshop (only for 3rd party filters that use it), Premiere Pro and After Effects.