AMD: R8xx Speculation

I don't know, but to me, showing off their DX11 silicon (albeit behind the curtain) certainly sounds like sending a clear message.
Of course there's still the option that the chip is actually bigger than 180 mm2 and can support a 256bit bus, but I'm not convinced that a GPU this size can replace RV770/RV790. GPUs are continually getting larger and nowadays, 200 mm2 is almost the value segment.

 

Well, they don't always get bigger. R600 -> Rv670 got smaller with similar performance. Same went for G70 -> G71. Granted those were just refreshes.

Still, if most of the required hardware for DX11 already existed in Rv770/790 then it's possible that even with theis small size they could increase performance a fair bit.

There's so much that we still don't know. Is it still basically the same arch? Did they radically change anything? Did they remove anything they now deem superfluous? Was there some breakthrough that has allowed for significantly higher clockspeeds?

Or perhaps there was just a general philosophy that being first to DX11 by a few months meant they would do the least possible to get something out ASAP while continuing work on a more revolutionary part?

Still Terry (CatalystMaker) seems overly excited about this chip, so there's gotta be something there (hopefully). I can see PR people overly hyping something, but I just don't see Terry needlessly getting this excited over just a nominal bump over Rv770/790.

Regards,
SB

 

This entire thing is pretty bizzare. ATi never published any slides, infos, die-shots etc. until launch. I smell... Dave...?

 

RV730 is 150mm2. RV740 is 136mm2. 14mm2 smaller and packing twice the SPU/TMU logic. Thats damn impressive to say the least, but here we are talking about a chip that is nearly 80mm2 smaller then RV770. A fair guess estimation would put Evergreen I would say 10%-30% faster then RV770. Not bad at all for it's size. This chip to me screams mainstream and the perfect replacement for the HD 4850/HD 4870/HD 4890. Faster performance, DX11, and and better margins then the former SKU's. Probably a good bit better margins actually.
I have a hunch though like many others, this is not the performance part. I want to say their is a 320mm2 part as well. Nothing wrong with a G92 size performance part.

 

Maybe they're picking up a CPU habit. Showing stuff running long in advance of availability.

 

That's not what I meant. R300, in its days, was a high-end GPU and its die area was around 200 mm2. Now look how cheap you can get a modern GPU that size today, and how big are the high-end dies.

 

Did I not scrub hard enough when showering this morning? I didn't think I'd be that pungent...

 

Margins depend on yields... now with the talk of ~20% for RV740, let's hope things improve drastically in the next 4-5months.
I agree, if this thing has a slight increase in shaders and the same clocks, if not higher, it would be a decent step up from the current ATi cards, I would say 10-30% faster than 4890.

320mm2 for the performance part seems a bit large, 1.8-2bil trannies(?). I was thinking something around 250-260mm2. Still that is based on the thought that Evergreen is more of a midrange replacement. I'm going to have take a look at some AMD/ATi slides and see how many segments they usually target.

 

Bison has this to say:

http://www.brightsideofnews.com/news/2009/6/3/ati-has-the-smallest-directx-11-chip.aspx

The whole article has a "it's smaller so it sucks" attitude comparing Evergreen to GT300 and puts Cypress as a x2 part versus GT300.

 

Is AMD planning to fight gt300 with a ~120mm2 die size saving? What do they think they have made here?

This entire thing is a great fud spreading smoke and mirrors strategy. There's something mighty fishing going on here. Working silicon, but no parts. No expected launch date and no perf/area/tranny/spec estimates either. Pure PR stunt.

I wouldn't bet on it being the rv770's successor. But then, if nv goes with tessellation-in-shaders, it may well turn out to be competitive.

 

Something fishy? Do you realize that RV770 is only 260mm2 right? If there is something fishy then it has been going for the last 12 months.

 

What? Does anyone thinks this 40nm wafer there is nothing more than a test vehicle in disguise...

 

By fishy, I mean that I don't think that this is the rv770/790's successor.

 

Im probably reading too much into the past here, but I don't think we should count too much on die size to indicate what units are on the chip. For instance, RV670 at 192mm^2 and RV770 at 265mm^2 went from 320 to 800 SP's - 2.5x the shaders for 38% more die size.

So if we compare across the same process, RV740 to this new die, even if its 137mm^2 -> 180ish mm^2, fitting 2.5x more shaders would be 1600 SP's. Now in all likelihood its less than that, and again, I'm probably reading too much into what ATI has done in the past, but it's certainly proof that we can't really use die size estimates to show that a chip will have those units.

 

Exactly. such miracle increases in density don't occur back to back often. The ALU's seem to be about ~25-40% larger for the chip shown (from rv770).

 

Last I heard AMD wasn't interested in competing directly with an enthusiast part like GT300. Rather they were competing in the segment just below that and then making an x2 part to maintain some presence in the enthusiast sector.

Regards,
SB

 

You heard correctly.

That's what I was comparing. Inefficiencies due to AFR included.

 

ATI with this show on computex showed 3 things:
- They are ready to jump on DX_11 with silicon and software working everything ready for Win 7 launch.
- They want to be the leaders in dx_11 and have developers working with them on dx_11.
- They confused everyone including nvidia showing the wafer. No one knows ATI strategy anymore so they are very unpredictable.

 

I haven't seen this posted anywhere, though it is old, so here it is.

 

Not posted before because it sounds like garbage.
Fudo and Theo have been so far off based so many times, espesially with up coming GPU's. Believe it or not, Charlie has actually been pretty good at it. Well ok... semi-accurate at least.