nVidia GTX 6xx series and AMD HD Radeon HD 7xxx series

TSMC and Global Foundries will be the Manufacturers of AMD HD 7000 series GPUs

TSMC and Global Foundries will be responsible for the manufacture of these new chips, though it is unclear which of them will manufacture models of low, medium or high ranges, although it could presumably TSMC will manufacture high-end chips ( and perhaps the mid-range), and Global Foundries may be responsible for manufacturing the low-end chips, which is not unreasonable. Considering that currently it is responsible for making the APU Llano, and they incorporate an IGP based on Radeon HD 6000 (VLIW-5D).

Nvidia Kepler vs AMD Radeon HD 7000 series

Using a 28nm manufacturing process means that the new Radeon HD 7000 series cards will have a lower consumption and higher operating frequencies margin, i.e., yield will be more even if they share the same design as the previous generation. But improvement is also expected in terms of the new architecture, so it will be an improvement for two reasons. Nvidia Kepler is no less than AMD Radeon HD 7000 series, as Nvidia has said that Kepler will be much faster than Fermi (with almost 2X performance).

Kepler promises to double the computing power of double-precision GPGPU compared to current Fermi, and to provide a performance / watt 50% higher than Fermi, Nvidia has not revealed details on its new architecture, but it is expected that as the Current Fermi (GeForce GTX 400 and 500 Series) is used in at least 1 or 2 families of products (Geforce GTX 600 and 700 Series) that will be available from late 2011(Release Date info Updated) until late 2013 when Maxwell may make his appearance.
Returning to Kepler, Nvidia has completed the design (known as Taped-Out), so no new features added with the exception of the base architecture optimizations.
It is expected that within a few months after Nvidia reveals some details on this new architecture.

One of the things that bugs me about AMDs launch is the use to two very different chips with the "same model number". First they have some chips use VLW4 while others use 5, then they rename the 5770 to the 6770. Now they are launching an entire series planned from the start to use two different arches. They are starting to act like another GPU designer I can mention.

The other thing that (possibly) bugs me is the 7870 being identical to the 6970. Only difference is 28nm vs 40, along with all the usual things that brings. I'm still trying to decide if this is good or bad.

i was actually hoping AMD would make a monolithic GPU design.

i still have a feeling that the same trend will happen for the past 3 years; the fastest AMD single GPU card is on par with the 2nd best single GPU from nvidia (4870:gtx260, 5870:gtx470, 6970:gtx570)

I don't see this as a bad thing. Performance/watt is very much in AMDs favor. By using CF they are able to compete with the high end. Yields are better for AMD because their chips are smaller. I think this strategy works very well for them.

I'd like to also point out the issue that Nvidia is having with scaling down to the low end. The 8800GT/9800GT is still faster then anything Nvidia has released to the low end in the last few years. The GT240, GT430, GT520, etc are all slower for games. The 8800GT was released how many years ago? If this keeps up Nvidia will have a very hard time making the lower end cards which is where the volume is.

The GTX 560 Ti was their first glimpse at 8800 GT value and performance. I bet that the GTX 660 will be a very nice 200 dollar card offering up great performance.

Heard the cards will only be released quarter 1 of 2012 as it went into production only a couple of days ago.

Question is do we really need faster Gpus or better designed games?

But how will they sell Gpus then.

Looking at the gaps in the different price ranges everything is filled up to 300. Are they going to fill that gap between the 6970/570 with the new Gpus and the gap between the 580 and 6990?

On the other hand there's intel with their new Igp with 3x monitor support and dx 11in the mix. Now that will be the thing everyone wants to see if they closed the gap and by how much.
Interesting times ahead. Get your popcorn.

good point. And what happens if the performance increase we see is not gpu itself but a benchmark on a cpu with a PCI-e 3.0 controller and the performance gain is actually PCI-e 3 due to next to nothing bandwidth? So we need to watch that as well. The overhead took something like 3Gb/s out some cards.

Thanks for opening this up Mal!
I have a question about the HD 7XXX series using Rambus VRAM. Isn't Rambus expensive? Wouldn't that drive up the prices to such a high extent? If it's supposedly 2x faster than GDDR5, wouldn't that performance reflect the price?

Same thing with Radeon HD drivers:

http://vr-zone.com/articles/amd-ra [...] 13766.html

Not really. Fermi was a disaster as launch. Overheating and noise were the major issues. The GTX 460 was the light in the dark tunnel that is Fermi. The GTX 500 series is Fermi improved. Cooler, quieter, but not a whole new arch. Not really a superior design compared to AMD's Northern Islands. However, Nvidia tends to have more stable drivers. The 6850 is certainly not crap. It performs very well and for a low price tag. It's practically on par with a GTX 460.

Thanks! I'm thinking of crossfire but i'm scared of the horror stories! ;O

Theres no horror stories. Just the 6870s in crossfire and the now and then Amd lucky packet. But always wait for the other blokes to update first. If they moan then you know keep the working ones lol

As it has been for well over a decade.

imagine if they become 1 what excellent Gpus we will get. But we will never be that lucky

Very true, I think you are on to something here.
We have not really had a solid table that compares generation to generation for a good while now. HD 2xxx was different to HD 3xxx the 4 series was about the same but then they stuck the 4770 and 4890 in to confuse things a bit 4-5 was similar then the 6 series was a total fiasco.
Seems like they may have settled for launching what used to be the mid life re spin as a new product for the lower speced cards while launching a new arch for the high end ?

Mactronix

But is that really an issue with the Fusion chips eating into the low end discrete card market. Nvidia have a decent Arch at the moment that rocks in the midrange, if they can do that again then I don't see an issue.
Older higher end cards being faster that new low end cards is nothing new.

Mactronix

Nice to see AMD using 2 fabs for the graphics chips. That should help with availability.

Another interesting fact is do anyone know why Amd used GDDR5 ram? The memory bus is still 256 bits wide, so the only way to get the extra bandwidth was to use faster GDDR5. ATI did this, going with a 1200 MHz GDDR5 for 153.6 GB/s, but it's a mere 33% improvement versus the 4870, and only 23% versus the 4890. Which is small when you compare the two chips twice the size 23% improvement.

Another thing is the Cypress has a onboard CP which is dma capable. Which means it can talk to the system ram without any help from the cpu.

They added two new shaders i.e the Domain and Hull shaders where with the previous gpu's tessellation was done via the vertex shader. They added a raster unit which was suppose to give it 1,7Tri/s. But they were clever in fooling everyone. Faster double scanning but your still sit with 1 triangle = 1 clock cycle like with the previous cards.

Nvidia cards can do 4 cycles per triangle

But the thing Amd wont give much details about is the dispatcher. Now that puppy is the scheduler. Everyone knows 4 threads equals 4 cycles gpu time. But if the threads are the same age pixel threads will be favored. Its also responsible for putting threads to sleep when they need to get resources thats in the system ram. Now it will put that threads to sleep until it spots a gap to do it. Now it almost works the same as a cpu would do with the hit and miss of its caches but a miss and the thread gets rescheduled sometimes downgraded. So it might not run after the one that replaced it and what do you people will happen when a piece of the puzzle is missing? And the cayman gpus doesnt have any support for over or underflow! What does it do when it happens signal a zero?

Its basically not the drivers but the design of Cayman itself especially with handling bad code. Amd gpus because of the registry and the scheduler are vulnerable to anything looking like a bad code.

The cayman was just a better tweaked version of the RV770 with a small increase here and there and 1 or two additions like the two shaders. scheduler and 16 ROPs. (Took em 4 years to increase it to 32 actually)

The new card will be just tweaked again and I'm sure the have found a better way for the scheduler to react to bad code other than throwing a big fat lie of a 0.

With the new rambus they got no option as they still use the same old 256k interface. They struggle to find a working design to increase that and to save dough so they just try to get faster ram to squeeze out that extra bandwidth. But one day nvidia is going to find a way to connect their bus in a more efficient manner and no ram XXXR or whatever is next will not help them.

Nvidia made the big fermi jump going from the "30 core" to the "32 core Gtx 480"
All the extra ram and bandwith were put there for DX 11 nothing else. take that advantage away from them against the older cards and you sit with a 40 performance increase with AA disabled. What did they promise use? 2x the performance than the old stock. Thats from both manufacturers. On Dx 9 and 10 the cards of today and probably future ones have a much bigger chance of bottlenecking a cpu due to the fact it clears the buffers faster than what the cpu can fill it. But thats also where they decided to supress sheer performance on dx 9 and dx 10 and went out to make the card able to handle dx 11.

Thats why we probably see the small increase on dx 9/10 and lower resolutions.

no your understanding it wrong. Do not think of ram like cpu ram need to save my game data. Cause that's wrong. The gpu got caches was increased to 64k on the new Gpus. But Amd know why they added a smaller bus on the 5770. First off all the card doesn't have what it takes to do dx 11 calculations. So it doesn't need the bandwidth. I'm talking about the 5870 cards and up that are able to handle Tesselation and beat they're pre accessory by twice the margin. The 5770 won't be able to do that because only that twice the performance gets thru with dx 11 where the older cards are pretty choked due to its using the vertex shader for Tesselation and the less bamdiwth its got. That's why we see the bigger ram amounts. Its for nothing else other than dx 11 there. Without that the twice the performance drop to 40 percent. So there's no use using a bigger bus on a gpu unable to do much dx 11 calculations. Amd was forced to use Gddr5 it wasn't the matter by choice. The wanted to use a bigger bus with cheaper ram. Unfortunately they couldn't get the design right.

Nvidia on the other hand were throwing away all they're performance advantage they manage to get over the previous cards by connecting their bus in a manner which make their gpu less efficient than it would have been.

If you look at the talk all around bout the cards it will still be the same design, faster ram to increase the bandwidth. But a thing except from shrinking it they wouldve tweaked are the scheduler and maybe get a bit more triangles per calculation out of the chip. I mean you can have superman as your ram your gpu will still not yield performance gains when it still use 1 clock cycle for a triangle calculation. Like I said nvidia use 4!!! Now that shows clearly how messed up that big old bus is connected.

Nvidia got rid of most of the caches and doesn't have a scheduler making calculated guesses like Amds one and that's one of the reason a nvidia gpu react better to a bad code.

Game developers don't know much about this scheduler. Not many people do. Amd doesn't give a good explanation for it or want to really discuss it. I respect that but how are you going to design a game when a piece of equipment that you know little about is going to be in charge of your work and decide which is a high and which is a low priority. I mean what happens to game when it gets a code returned 0 or nand or whatever they use when there a bit of bad code?

The only way Amd is going to get major performances is if they increase the cycles per triangle and get rid of that scheduler.

For nvidia. Not hundreds of connections to your memory. Short and sweet like Amd.

How many people knew that you get Tesselation in dx 10 as well? How many people knew Amd Gpus could do Tesselation before Cayman?

The Rv7 chip has very little difference to Cayman when you go really inspect under the hood. More tweak here and there. Its not a completely new design.

nice explanation of Tesselation but what does it mean? Really. That's from a designers point of view. That's what Tesselation is suppose to do and can do. Theoretical. Now they added a extra raster so your gpu is suppose to run over 1.7GTris/s. But it runs at just over 800. That'sthe theoretical cut in half in reality.

That's the Tesselation performance. Why is it running half the theoretical? Because the gpu can do only 1 triangle per cycle. If it could do 4 you'll be looking closer to 1.7.

And Vram doesn't tell you all the bandwidth on the gpu. What about the L2 which is the same size as with the old stock? Another 5.8mb or registers and add the fact that the byte per pixel size is increasing by a big amount and its going to keep on growing at every gpu release. The Vram bandwidth doesn't matter how big doesn't tell you the bandwidth of the L2 does it. What happens when the gpu needs to add get a lot of fractions or try to avoid shader stalls? That poor old L2 cache will suffer for bandwidth.

Its like a cpu. Put 20gb ram in it but give it a 64k cache. Will that huge amounts of ram be any use when it has to run around in the ram to get the data all the time?

The question I have to ask is... Get a MSI GTX 570 Hawk OC'd in Feb or wait till June for an equivalent in the Keplar range price wise... ?

AMD will get to market their products earlier. and considering that they now have 2 fabs to work on it is possible.

NVIDIA on the other hand will have probably release their big guns earlier or as early as AMD, 680/670 as most of the time their top tiers share the same board / chips.

but a simple rule of thumb exists, if your budget conscious and is targeting a top tier GEFORCE card don't be an early adopter.

the old GTX260 192 still holds the record for the fastest CRAPIFIED card ever. the 480 (not much the 470) has a very distinct after taste as well.

More performance is as far as cores and memory bandwidth is great for FPS
but what AMD needs to do is to get more developer support for ATI Stream (or whatever they are calling it now)
In the workstation market Nvidia is dominant primarily because of the support for CUDA in almost all of the software.
Unless the software devs go to an open standard like OpenCL or start using DirectCompute API then Nvidia will maintain their edge in the enterprise market
with so much support for CUDA
It will not only benefit the end consumer but the video card makers themselves
if an open standard for GPGPU computing is adopted.
(though some could say that Nvidia wouldnt support this)

Imagine when most of your apps could take advantage of GPGPU processing
I know that by using ATI Stream for my encoding/rendering has reduced times
drastically for me
Software devs would especially benefit if there was an open standard adopted

It is nice to see that the new AMD arch is designed to be better at GPGPU
but I hope software devs can catch up