[hkultala]

Originally Posted by TKK

You could even say 6xxx encompasses four architectures, or at least architecture revisions:

...

- Evergreen + UVD3 (Llano)
- Northern Islands VLIW5 / Enhanced Evergreen (Barts, Turks, Caicos)

Any basis for the claim that these are different? (when not counting the memory architecture which is of course different because there is also cache-coherent CPU in the same chip)

So yeah, I agree that HD 7000 for Trinity is just for marketing, APUs will always lag behind discrete GPUs by at least one generation.

We don't even know for sure how many of the GPU's that ATI is going to release in the coming winter are "GCN" or just cayman's VLIW4 architecture manufactured with 28nm.

And I would not say "always lag", Llano had their first GPU manufactured with "CPU process", and Trinity will be only second, it also takes time to "develop the development process" so that they can quickly integrate their new GPU's into APU's.

(bobcat-based chips use "gpu processes")

[CarstenS]

Originally Posted by hkultala

We don't even know for sure how many of the GPU's that ATI is going to release in the coming winter are "GCN" or just cayman's VLIW4 architecture manufactured with 28nm.

I'd guess 0 of 0. ATI isn't going to release GPUs anymore, AMD is. *SCNR*

[TKK]

Originally Posted by hkultala

Any basis for the claim that these are different? (when not counting the memory architecture which is of course different because there is also cache-coherent CPU in the same chip)

Granted, I don't know for sure.

That being said, my reasoning goes like this:
To keep the development as simple and short as possible, it would be easiest to take an existing GPU design and fuse it with the CPU. The only 5-SIMD design AMD made and probably had ready when Llano was designed was Redwood, and I assume Llano's GPU is based on it.
The UVD engine is probably easy to upgrade since it's a bit separated (layout-wise) from the rest of the GPU. But Barts & Co. had their TMUs and Tesselation unit updated as well.
I'm not saying it's impossible that AMD ported those over to Llano, but considering the market & performance segment Llano is targeting, I think it's at least possible that they didn't bother. They had their hands full enough with solving more pressing issues like yields and power consumption, I think.

But you're right, it might me architecturally identical to NI-VLIW5 GPUs.

Originally Posted by hkultala

And I would not say "always lag", Llano had their first GPU manufactured with "CPU process", and Trinity will be only second, it also takes time to "develop the development process" so that they can quickly integrate their new GPU's into APU's.

Fair enough, they may get there some time in the future. The time gap will certainly shrink, but I think it's easier (and less risky) to integrate a finished design than a work-in-progress design, that's why I wouldn't expect, let's say, Trinity successor with GCNv2 two months after discrete GCNv2. Unless they're made on the same manufacturing process, then it might be a somewhat different matter.

[Psycho]

Ivy Bridge GPU slides from IDF: http://www.anandtech.com/Gallery/Album/1375#1

[fellix]

Originally Posted by Anand

Intel's Mooly Eden just disclosed Ivy Bridge's transistor count to a captive audience at IDF this morning: 1.48 billion transistors. That's presumably for the quad-core version compared to 995M in Sandy Bridge. That's nearly a 50% increase in transistor count. Don't pay attention to the die shot above, that's not an accurate representation of Ivy Bridge. Intel is holding off on showing the die until closer to launch. Why? Likely to avoid early disclosure of how much of the die is dedicated to the GPU. As you'll see in our Ivy Bridge architecture piece later today, the lion's share of those transistors are improvements in the GPU.

Damn! Intel is serious on the IGP front here.

[pjbliverpool]

wow an extra half a billion transistors for the GPU alone is pretty serious indeed. We may finally start seeing good intergrated PC graphics across the board.

[fellix]

Let's hope the driver support will be on the same level as the HW investment. Carmack for sure will be pleased this time.

[hoho]

Originally Posted by pjbliverpool

We may finally start seeing good intergrated PC graphics across the board.

Too bad they have to get by with mediocre memory bandwidth.

[Kaotik]

Originally Posted by pjbliverpool

wow an extra half a billion transistors for the GPU alone is pretty serious indeed. We may finally start seeing good intergrated PC graphics across the board.

You seriously suggest that only GPU is upgraded in IB compared to SB?

[pjbliverpool]

Originally Posted by Kaotik

You seriously suggest that onnly GPU is upgraded in IB compared to SB?

Its anand suggesting those transistors are mostly attributed to he gpu nkt me. I certainly hope the cpu has some decent improvents over SB.

[DavidC]

Originally Posted by pjbliverpool

Its anand suggesting those transistors are mostly attributed to he gpu nkt me. I certainly hope the cpu has some decent improvents over SB.

It's a Tick so, the CPU may see 3-4% gains if at all. Even then, CPU and I/O transistor counts are always little compared to others. For comparison, there was zero performance gain for Westmere, for existing software. Interesting there is an L3 cache on the GPU now. The shared L3 cache is called LLC.

[Alexko]

Originally Posted by pjbliverpool

wow an extra half a billion transistors for the GPU alone is pretty serious indeed. We may finally start seeing good intergrated PC graphics across the board.

Isn't there some extra cache as well?

[DavidC]

Originally Posted by Alexko

Isn't there some extra cache as well?

Nope.

[Kaotik]

Originally Posted by pjbliverpool

Its anand suggesting those transistors are mostly attributed to he gpu nkt me. I certainly hope the cpu has some decent improvents over SB.

Ah indeed, I missed the last phrase there - however there's now update on the article - Intel gave real numbers, SB 1.16B, IB 1.4B, so only ~20% bigger transistorbudget, which of only part (even if it's lions share like Anand said) for GPU

[mczak]

anand updated that transistor count quote: http://www.anandtech.com/show/4798/i...8b-transistors - apparently intel doesn't always count them the same way...

Still, assuming most of the increase is due to the gpu (and I can't see why not) it's roughly ~200 million transistors more for the gpu. That certainly looks like a quite big increase (though I don't remember having seen any number for transistor count of the gpu of SNB alone). And certainly it's not just cache - I see no good reason for the L3 gpu cache to be large (interestingly, on the slides intel actually doesn't mention L3 gpu cache is there for performance reasons, just for lower power consumption...).

Looks indeed like the gpu part could be quite competitive with Llano - it's got the features (3 displays, d3d11), it might still not quite catch the fastest Llano versions but could at least be close probably. Now of course Trinity is a different matter, it should be faster but will appear (how much really?) later.

[Kaotik]

Originally Posted by mczak

anand updated that transistor count quote: http://www.anandtech.com/show/4798/i...8b-transistors - apparently intel doesn't always count them the same way...

Still, assuming most of the increase is due to the gpu (and I can't see why not) it's roughly ~200 million transistors more for the gpu. That certainly looks like a quite big increase (though I don't remember having seen any number for transistor count of the gpu of SNB alone). And certainly it's not just cache - I see no good reason for the L3 gpu cache to be large (interestingly, on the slides intel actually doesn't mention L3 gpu cache is there for performance reasons, just for lower power consumption...).

Looks indeed like the gpu part could be quite competitive with Llano - it's got the features (3 displays, d3d11), it might still not quite catch the fastest Llano versions but could at least be close probably. Now of course Trinity is a different matter, it should be faster but will appear (how much really?) later.

You'd need around 3 times the performance of their current GPU to even match Llano


Going by the trend the graphs (http://www.techpowerup.com/reviews/A...rmance/13.html) FullHD would just grow the difference in Llanos favor.

[mczak]

Originally Posted by hoho

Too bad they have to get by with mediocre memory bandwidth.

It's not really that bad, considering you can still get midrange cards like the GTS 450 or HD6670 with 128bit ddr3 memory (granted you shouldn't get them...). Of course compared to the "real" cards which have gddr5 memory it's not quite half the bandwidth, but the ratio of flops/bandwidth is still roughly similar (well in intel's case they actually don't have that many flops).
Also IGPs can use the LLC cache which potentially saves quite a bit of bandwidth. And things like dynamically generated vertex data would never need to hit memory (or the pcie bus for that matter).
Memory bandwidth won't scale that well though for future chips.
Now SNB-E would have twice the bandwidth too bad it won't have a IGP .

[mczak]

Originally Posted by Kaotik

You'd need around 3 times the performance of their current GPU to even match Llano

Going by the trend the graphs (http://www.techpowerup.com/reviews/A...rmance/13.html) FullHD would just grow the difference in Llanos favor.

I was going by these results where the difference is more like a factor of 2:
http://ht4u.net/reviews/2011/amd_lla...a8/index16.php
I guess (apart from different titles) the average is better because these tests used lowest settings, plus lower resolutions (indeed shows worse scaling for HD3000 there too).
But you're right that Ivy Bridge won't be able to touch A8-3850. But it might be close to the lower-end models, and things should be (because I expect a larger clock discrepancy for Llano than for Ivy Bridge for mobile parts) closer on the mobile side a bit.

[Paran]

Originally Posted by Kaotik

You'd need around 3 times the performance of their current GPU to even match Llano

If we compare the top models its about two times. I don't see it coming though, latest rumours hinted 60% in Vantage.

Anand was wrong with his 50% increase. Sandy bridge has 1.16 Billion transistors while Ivy Bridge 1.4 Billion. This is a ~20% increase in transistor count.

[ToTTenTranz]

Originally Posted by Kaotik

Going by the trend the graphs (http://www.techpowerup.com/reviews/A...rmance/13.html) FullHD would just grow the difference in Llanos favor.

Looking at the difference between 1680*1050 and 1280*1024, I'd say the difference is actually narrowing.

Let's not forget Llano's GPU is severely bandwidth limited, which shows in higher resolutions, whereas HD3000's access to L3 pays off.

[DarthShader]

Originally Posted by pjbliverpool

wow an extra half a billion transistors for the GPU alone is pretty serious indeed. We may finally start seeing good intergrated PC graphics across the board.

Well, it's only half of half billion after all and most of these will be to bring DirectX11 and OpenCL functionality to the GPU, not performance. Have you forgotten how it was with HD5000 series? Much more transistors, not much more raw performance.

[mczak]

Originally Posted by DarthShader

Well, it's only half of half billion after all and most of these will be to bring DirectX11 and OpenCL functionality to the GPU, not performance. Have you forgotten how it was with HD5000 series? Much more transistors, not much more raw performance.

I don't think that's really "most" of the transistors being used for DX11 - RV710 was 242 million transistors, whereas Cedar was 292 million transistors. Granted SNB (HD3000) GPU is faster than that (so maybe a bit more transistors) but certainly Ivy Bridge doesn't need ~200 million transistors just for some dx11 features...

[Erinyes]

Originally Posted by DarthShader

Well, it's only half of half billion after all and most of these will be to bring DirectX11 and OpenCL functionality to the GPU, not performance. Have you forgotten how it was with HD5000 series? Much more transistors, not much more raw performance.

Anand has since corrected his numbers, the Transistor count is 1.16 billion for SB v/s 1.4 billion for IB, which is about 20% more. Most of these would be towards the GPU but there are some improvements to the CPU cores as well (probably a Conroe to Wolfdale type change). Number of EU's remain the same(16) so the performance increase is only due to arch improvements and higher clocks (say 20%). Overall Vantage score increases by about 60% according to Anand, and 3DMark 06 by about 30%.

And anyway, HD5870 brought a ~50-60% performance increase over Hd4890. I would hardly classify that as "not much"

Originally Posted by mczak

I don't think that's really "most" of the transistors being used for DX11 - RV710 was 242 million transistors, whereas Cedar was 292 million transistors. Granted SNB (HD3000) GPU is faster than that (so maybe a bit more transistors) but certainly Ivy Bridge doesn't need ~200 million transistors just for some dx11 features...

Well most of the transistors would be for the GPU. Apart from DX11 they are also beefing up the media capabilities (three displays, faster quicksync, etc). While the CPU side is also going to be improved somewhat, it isnt anything like the changes to the GPU (you could say CPU is a tick and GPU is a tick+)


Edit: Also with regard to Trinity, it looks like we could be heading towards an early 2012 launch, so it might beat Ivy Bridge!

[CarstenS]

Originally Posted by Kaotik

You'd need around 3 times the performance of their current GPU to even match Llano


Going by the trend the graphs (http://www.techpowerup.com/reviews/A...rmance/13.html) FullHD would just grow the difference in Llanos favor.

I disagree. Going from 1280 to 1680 actually lessens the gap when comparing the ratings in that review.

Speaking of which: How much sense does a rating make when incorporating DX11-tests where one competitor fails completely because he lacks DX11 support, while the other has it, but struggles to even achieve two-digit Fps, making the game completely unplayable. One would rather think, it's make sense to at least tune the settings so as to at least one of the competitors can achieve playable Fps.

[Paran]

Originally Posted by Erinyes

Number of EU's remain the same(16) so the performance increase is only due to arch improvements and higher clocks (say 20%).

EUs increased. SB has 12 EUs while Ivy has 16 EUs (GT2 part).

Originally Posted by Erinyes

Edit: Also with regard to Trinity, it looks like we could be heading towards an early 2012 launch, so it might beat Ivy Bridge!

Unlikely. Trinity production starts early in 2012 which makes a Q2 launch likelier. Not much difference between Ivy and Trinity, maybe Ivy Quad will come 1-2 months earlier.