Nvidia GPUs soon a fading memory?

[Lonbjerg]

Basically AMD shot themselves in the foot here.
Their HyperTransport bus is so good that there is little to gain by making single-die 'native' solutions.

Yield wise it's better with a GPU core and a CPU core too...but I don't get all the attention to a mainstream CPU with and "Integrated Graphics Processor"...gamers will still want a add-in board, and thus wasting a major part of the core.

 

[Skurge]

Yield wise it's better with a GPU core and a CPU core too...but I don't get all the attention to a mainstream CPU with and "Integrated Graphics Processor"...gamers will still want a add-in board, and thus wasting a major part of the core.

Llano is mainly aimed at OEMs and people who buy low end GPUs

 

[Scali]

Yield wise it's better with a GPU core and a CPU core too...but I don't get all the attention to a mainstream CPU with and "Integrated Graphics Processor"...gamers will still want a add-in board, and thus wasting a major part of the core.

I think that is a good observation. Fusion will probably fall in 'no mans land'. Faster, more power-hungry (and more expensive) than a traditional super low-end GPU, but not fast enough as a proper discrete card, so still not an interesting solution for gaming, GPGPU etc.

Perhaps there is no market for it. Office machines, servers and most notebooks will do fine with Intel's ultra low-end IGPs. I think HTPCs can also work with IGPs these days. I've seen Atom+Ion systems handling BluRay content just fine. I think Intel's GMA HD can do it too.

 

[taltamir]

AMD will be fine because they ditched hector ruin.
nVidia will be fine because they are shifting to new markets and having great success in those. (fermi while too hot, too big, etc as a GPU; is a huge success in the computing market, where nobody can even touch its performance)

 

[Scali]

AMD will be fine because they ditched hector ruin.
nVidia will be fine because they are shifting to new markets and having great success in those. (fermi while too hot, too big, etc as a GPU; is a huge success in the computing market, where nobody can even touch its performance)

I think the GTX460-series will be the most important test for Fermi.
Those should be the mainstream/volume parts of the Fermi architecture. If nVidia prices them right, they could be a success.

 

[taltamir]

indeed, it is all about price in the end.
and AMD and nvidia have no reason to bring it down.

 

[Skurge]

I think the GTX460-series will be the most important test for Fermi.
Those should be the mainstream/volume parts of the Fermi architecture. If nVidia prices them right, they could be a success.

The way AMD and Nv are going, they will have priced to fill in the gaps AMD conveniently left in their line-up and prices wont shift at all.

Makes you wonder :hmm:

 

[taltamir]

The way AMD and Nv are going, they will have priced to fill in the gaps AMD conveniently left in their line-up and prices wont shift at all.

Makes you wonder :hmm:

price collaboration between opponents might be illegal, but it happens all the time.
And as long as they don't formally conspire to do so, but merely do so without talking to each other it is actually not illegal.

nVidia has been doing it for a while, and AMD is the one who has been pushing the price war, it seems now that AMD wizened up and realized what nvidia is trying to do and started playing along, for the benefits of both (but not the benefit of us)

 

[GaiaHunter]

but I don't get all the attention to a mainstream CPU with and "Integrated Graphics Processor"...gamers will still want a add-in board, and thus wasting a major part of the core.

The interest that Llano is picking is two-fold:

1st) We've seen that BD architectures seems to be taking a bigger emphasis in the Integer operations than FP. Off-loading some/most of those to the core of a GPU companion might be the path AMD is taking;

2nd) Gaming market. While PCs equipped with reasonable discrete GPUs are much powerful (and much more expensive) than current generation consoles, the fact is the majority of PCs out there uses Intel graphics or other IGPs. So for what market will devs program?

The rumours that put the graphic power of Llano in the 5570 realm, quite weak compared to mid and high-end graphic cards but so much more powerful than IGPs, if true can have an impact in PC gaming market. No longer will those OEM boxes ppl buy for $500 in whatevershitshop be completely useless for gaming. And once people taste decent gaming in a PC they might be enticed to upgrade their experience, instead of having a horrible experience with an IGP.

As you can see in responses to this thread, it is too good to be true. But because of that, even if we take these rumours with a grain of salt (or a truck load of it), it does pick interest.

 

[Scali]

1st) We've seen that BD architectures seems to be taking a bigger emphasis in the Integer operations than FP. Off-loading some/most of those to the core of a GPU companion might be the path AMD is taking;

But Llano is not a BD architecture, nor is there any effort to off-load anything to the GPU. Heck, you can't even offload anything to a discrete AMD card at the moment, because AMD hasn't sorted out OpenCL yet... and there are no applications using OpenCL (partly because AMD is actively blocking OpenCL... which works against them, since companies such as Adobe just back Cuda).

2nd) Gaming market. While PCs equipped with reasonable discrete GPUs are much powerful (and much more expensive) than current generation consoles, the fact is the majority of PCs out there uses Intel graphics or other IGPs. So for what market will devs program?

I don't know if IGPs are the reason why console gaming is so popular.
Given the cost of a decent discrete card, I don't really think that's where the bottleneck for PC gaming is (a real Radeon 5570 can be had for under $100, and the Radeon 5770 is a 'grown up' gaming card for less than $200).

 

[taltamir]

I don't know why people think intel IGPs are useless, they have stable drivers, more then enough capability to accelerate window, and are the most power efficient GPUs on the market.
If I need a GPU for a computer that doesn't run games, intel IGPs are an excellent choice.

Now, when it comes to gaming performance, intel IGPs are crap (not that nvidia or AMD IGPs are much better; 3 times zero is still zero)

 

[GaiaHunter]

But Llano is not a BD architecture, nor is there any effort to off-load anything to the GPU. Heck, you can't even offload anything to a discrete AMD card at the moment, because AMD hasn't sorted out OpenCL yet... and there are no applications using OpenCL (partly because AMD is actively blocking OpenCL... which works against them, since companies such as Adobe just back Cuda).

I'm not saying Llano is the thing - Llano is only the first step.

Yes, AMD will have to make sure the tools are in place, read collaborate with devs.

Also, while the Llano cores are still old K10 architecture, the uncore portion might be closer to bobcat (complete unsubstantiated speculation).

I don't know if IGPs are the reason why console gaming is so popular.
Given the cost of a decent discrete card, I don't really think that's where the bottleneck for PC gaming is (a real Radeon 5570 can be had for under $100, and the Radeon 5770 is a 'grown up' gaming card for less than $200).

I'm not saying console is popular because of IGPs - I'm saying PC gaming might not be as popular because of it, which is different.

For a person that doesn't keep in contact with the hardware world and then go out and buy a $500 OEM machine based on the size of the HDD or memory, it might not be so intuitive that something a discrete card can make a world of difference.

So they go home install a game and see a slow game with really crappy graphics that worse than any console, when they have a console and the games look better and are so much faster and rewarding.

How many will blacklist PC gaming just there?

Then even those that will think of buying a discrete card are faced with some barriers - there are discrete cards that cost $50 and those that cost $150 and even those that cost $500+! Which one should they buy? Is paying 3x or 10x worth it? Is paying an extra 20% or 100% of the total cost of your system worth?

Can they install it? Will the warranty be void if they open the box?

Will it actually fit in the box because they just bought that cute tiny PC but now their graphic card is massive in comparison?

Will their PSU have enough power to supply it? Will it even have enough connectors? Another $50-100 in PSU?

So that PC that cost $500 is now $700+?

For those that doesn't know can be a nightmare experience.

Or they could just buy it for the same $500 and have something like a 5570 that will make their games somewhat comparable to their console games. Maybe PC gaming isn't that bad...

----------------------------------------------------------------------------

I know you don't think it is possible for Llano to have GPU power around 5570 levels with only the available bandwidth.

Opposed to some in here I don't think AMD will be using sideports, eDRAM or extra memory bandwidth + memory on motherboard.

So 128-bits DDR3. BD fusion will have 256-bits, but that is not here and now.

As we have seen 25.6 GB/s isn't that bad. It is only bad when the CPU core access the memory heavily at the same time the GPU has to fetch textures or other maps, vertices or write frame buffer data.

In that case performance will be crippled.

What we need to know is fetch/write bandwidth over time both for CPU and GPU.

Possibly much easier than all those previous ways to increase bandwidth, would be increasing the size of the already existent caches in the GPU architecture, possibly the L2$, which are the ones coupled to the memory controllers (something that is not possibly to do with current IGPs, since they aren't directly connected to the memory controller) and responsible for memory operations.

Another interesting thing to see is if AMD is bringing back hybrid crossfire, allowing those extra SPs to be used by even high-end discrete cards and what kind of performance gains/power usage gains (if any), will that bring.

 

[Scali]

Yes, AMD will have to make sure the tools are in place, read collaborate with devs.

Then Fusion is doomed from the start.
Don't get me started on AMD devrel...

Or they could just buy it for the same $500 and have something like a 5570 that will make their games somewhat comparable to their console games. Maybe PC gaming isn't that bad...

I think that theory fails for the same reasons that discrete cards fail: How will they know that they need Llano?
9 out of 10 PCs in the store will NOT be Llano, because about 7 of them are Intel, and of the remaining AMD systems, two of them will be other types of CPUs, either with IGP or with a discrete card.

What we need to know is fetch/write bandwidth over time both for CPU and GPU.

Multithreading will be the death of Fusion.
If a game developer wants to optimize the engine by doing background processing on the spare cores while one core handles graphics, there goes your bandwidth.

(something that is not possibly to do with current IGPs, since they aren't directly connected to the memory controller)

I don't think it'd be very useful. These caches are far slower than GPU caches.
The cache in a Phenom II has only about 26 GB/s bandwidth(!). Most GPUs have faster memory than that.

 

[GaiaHunter]

Then Fusion is doomed from the start.
Don't get me started on AMD devrel...

Well then AMD might just close their doors, don't you think?

They'll never improve or learn...

I think that theory fails for the same reasons that discrete cards fail: How will they know that they need Llano?
9 out of 10 PCs in the store will NOT be Llano, because about 7 of them are Intel, and of the remaining AMD systems, two of them will be other types of CPUs, either with IGP or with a discrete card.

They don't.

They will read 4 cores - $500. And then they'll read 2 cores $600.



Additionally AMD at least thinks they can even out some of the playing market with the settlement they got with Intel.

If they can or not, is beyond me.

Multithreading will be the death of Fusion.
If a game developer wants to optimize the engine by doing background processing on the spare cores while one core handles graphics, there goes your bandwidth.

I don't think it'd be very useful. These caches are far slower than GPU caches.
The cache in a Phenom II has only about 26 GB/s bandwidth(!). Most GPUs have faster memory than that.

I'm talking about the GPU cache, not CPU cache (yes, ATI GPUs have cache).

 

[evolucion8]

I don't think it'd be very useful. These caches are far slower than GPU caches. The cache in a Phenom II has only about 26 GB/s bandwidth(!). Most GPUs have faster memory than that.

Do you have proof of that, because doesn't seems reasonable that the Cache bandwidth is only twice faster than it's memory controller. My CPU can do up to 63GB/s in Read/Write speeds and up to 120GB/s in copy operations. It pales in comparison for example, with current GPU's that have caches running beyond 480GB/s.

 

[GaiaHunter]

Do you have proof of that, because doesn't seems reasonable that the Cache bandwidth is only twice faster than it's memory controller. My CPU can do up to 63GB/s in Read/Write speeds and up to 120GB/s in copy operations. It pales in comparison for example, with current GPU's that have caches running beyond 480GB/s.

And the cache I was talking that is linked to the MC is GPU not CPU as Scali assumed.

 

[Scali]

Do you have proof of that, because doesn't seems reasonable that the Cache bandwidth is only twice faster than it's memory controller. My CPU can do up to 63GB/s in Read/Write speeds and up to 120GB/s in copy operations. It pales in comparison for example, with current GPU's that have caches running beyond 480GB/s.

See Everest L2 cache benchmark:
http://www.legionhardware.com/articles_pages/amd_phenom_ii_x6_1090t_be_1055t,2.html

 

[Martimus]

Do you have proof of that, because doesn't seems reasonable that the Cache bandwidth is only twice faster than it's memory controller. My CPU can do up to 63GB/s in Read/Write speeds and up to 120GB/s in copy operations. It pales in comparison for example, with current GPU's that have caches running beyond 480GB/s.

Phenom L3 Cache has around the same bandwidth as Dual channel DDR3, although the latency is about 1/5th the time. Phenom L1 and L2 cache is much quicker on the other hand.

This was a Phenom II 940. I didn't see a newer one on the web, but I have seen them before so they are available.

 

[Scali]

Well then AMD might just close their doors, don't you think?

They'll never improve or learn...

I've been developing using ATi products since the Radeon 8500... and no, they haven't learnt anything so far.

They don't.

They will read 4 cores - $500. And then they'll read 2 cores $600.

Yea right, as if AMD is going to give away the GPU for free, and throw in some extra cores while they're at it.
At the very least, Llano will be the same price or more expensive than current quadcore models from AMD, with or without IGP.

Additionally AMD at least thinks they can even out some of the playing market with the settlement they got with Intel.

If they ever get any good products on the market, perhaps...
But as I said, Llano is likely to fall in 'no mans land'.
Intel's integrated GPU makes a lot of sense from a business point of view. It's very low end, but serves regular office and non-gaming end-users just fine. They simply reduce the overall cost of the system.
AMD is (again) designing a chip that is quite large and complex, which isn't very good for yields, and hence cost... and it probably won't be able to really bridge the gap between IGPs that are useless for gaming, and discrete cards that are more or less game-worthy.
In that case, it provides no additional value over super low-end GPUs from Intel.

I'd like to point back to taltamir's post a while ago... he was right on the money:

I don't know why people think intel IGPs are useless, they have stable drivers, more then enough capability to accelerate window, and are the most power efficient GPUs on the market.
If I need a GPU for a computer that doesn't run games, intel IGPs are an excellent choice.

Now, when it comes to gaming performance, intel IGPs are crap (not that nvidia or AMD IGPs are much better; 3 times zero is still zero)

 

[GaiaHunter]

Yea right, as if AMD is going to give away the GPU for free, and throw in some extra cores while they're at it.
At the very least, Llano will be the same price or more expensive than current quadcore models from AMD, with or without IGP.

Isn't it exactly how they are competing atm? 4 cores for the price of 2 intel ones and 6 for the price of 4 intel ones?

If they ever get any good products on the market, perhaps...

OK.

Didn't you say some posts ago that you use some AMD products?

What the hell are you doing buying bad products?

But as I said, Llano is likely to fall in 'no mans land'.
Intel's integrated GPU makes a lot of sense from a business point of view. It's very low end, but serves regular office and non-gaming end-users just fine. They simply reduce the overall cost of the system.
AMD is (again) designing a chip that is quite large and complex, which isn't very good for yields, and hence cost...

This huge, low yield chip you are talking about is estimated around 170mm^2.

and it probably won't be able to really bridge the gap between IGPs that are useless for gaming, and discrete cards that are more or less game-worthy.
In that case, it provides no additional value over super low-end GPUs from Intel.

And that is the main question.

Of course if it can bridge the gap and allow reasonable gaming the game changes. Especially singe GPGPU is going to have an impact in PC applications and aren't Intel and NVIDIA betting that it will?

I'd like to point back to taltamir's post a while ago... he was right on the money:

Because normal users also need i7 to run windows, use voip, run ie and office since Athlon II can't...

 

[happy medium]

Kinda funny, there is the same thread in the Semi accurate forum.

http://www.semiaccurate.com/forums/showthread.php?t=2481

 

[GaiaHunter]

Kinda funny, there is the same thread in the Semi accurate forum.

http://www.semiaccurate.com/forums/showthread.php?t=2481

This is thread has long since moved from any possible demise of NVIDIA into something else.

 

[Scali]

Isn't it exactly how they are competing atm? 4 cores for the price of 2 intel ones and 6 for the price of 4 intel ones?

Note that I said AMD quadcores... AMD competes against its own current product line here.
Aside from that, I think AMD's current strategy is suicide.

Didn't you say some posts ago that you use some AMD products?

I was talking about CPUs here (the topic was Intel and the settlement, remember?).
I haven't bought an AMD CPU in years.

This huge, low yield chip you are talking about is estimated around 170mm^2.

It's all relative, isn't it?

And that is the main question.

Not for me.

Of course if it can bridge the gap and allow reasonable gaming the game changes. Especially singe GPGPU is going to have an impact in PC applications and aren't Intel and NVIDIA betting that it will?

nVidia has actually done something about it.
AMD's GPGPU currently is non-existent. No developer ever used their Brook+/Stream SDK because it was utter crap. And OpenCL support isn't available to AMD end-users yet (and AMD is holding OpenCL-supporting projects such as Bullet Physics hostage with nasty NDAs).

Because normal users also need i7 to run windows, use voip, run ie and office since Athlon II can't...

What does an i7 have to do with anything?
He's just saying that Intel IGPs get the job done, at minimum cost and minimum power consumption. For a lot of people, that is the only thing they're interested in, for their graphics chip. And in many cases, the OEM makes the choices for them. Lots of laptops have Intel IGPs for example. Most standard office machines do aswell (Dell, HP etc).
Huge HUGE markets that Llano won't be able to tap into.

 

[taltamir]

Try to avoid overly elitist "the average consumer is so damn stupid, they can't even tell what a GPU is" please. I have seen many companies brought down by taking this approach. And then they find out that people DO know how to use google and can comprehend basic things.

Sure my grandma doesn't know what a video card is, but my grandma doesn't play video games.

An IGP, whether integrated into the CPU or the northbridge, is wholly unsuited for playing games. It needs to be simple, cheap, and power-efficient. Its one and only task is to run windows.
And explaining to someone that they need a discreet video card to play games is simple enough that even the stupid can comprehend it.

What I Would like to see is desktop optimus, and something similar from AMD. This would be very useful feature which is being completely ignored by everyone (but nvidia and AMD promised it YEARS ago and neither delivered).

 

[GaiaHunter]

Note that I said AMD quadcores... AMD competes against its own current product line here.
Aside from that, I think AMD's current strategy is suicide.

Hmm and gets nothing else if it gets decent CPU+GPU package...

I was talking about CPUs here (the topic was Intel and the settlement, remember?).

And ATi isn't a brand of AMD? Quite convenient when talking about a CPU+GPU product to simply point the downsides of AMD as a CPU maker and forgetting the AMD upsides as GPU maker.

I haven't bought an AMD CPU in years.

And did you buy one in Athlon X2 vs Pentium D era?

It's all relative, isn't it?

Basically it is the same size of current Athlons X4 but with a GPU.

nVidia has actually done something about it.
AMD's GPGPU currently is non-existent. No developer ever used their Brook+/Stream SDK because it was utter crap. And OpenCL support isn't available to AMD end-users yet (and AMD is holding OpenCL-supporting projects such as Bullet Physics hostage with nasty NDAs).

So now that they are doing you complain?

http://developer.amd.com/gpu_assets...ent and Future with ATI Stream Technology.pdf

What does an i7 have to do with anything?
He's just saying that Intel IGPs get the job done, at minimum cost and minimum power consumption. For a lot of people, that is the only thing they're interested in, for their graphics chip. And in many cases, the OEM makes the choices for them. Lots of laptops have Intel IGPs for example. Most standard office machines do aswell (Dell, HP etc).
Huge HUGE markets that Llano won't be able to tap into.

Whenever people are given more resources they will tap into them and use them.

People don't do more with their IGPs because it can't do more. Give people more from their "IGPs" and they will do more.