AMD: R9xx Speculation

What do you mean by "even when tessellating"? Right now triangle throughput on Evergreen is faster with tessellation off than on. Is that going to change?

 

Yes, it is possible.

 

Ah you seem to know more than I do .
In any case that wouldn't be too bad neither. If you look at Tridam's numbers the maximum he got on the GTX460 was 750MTris/s with tesselation - and that only with "high culling" (radeons did not show any effect on culling with triangle rate). So 4 tris every 5 clocks at 900Mhz would be 720MTris/s - which is just slightly less and if Barts continues to not show dependency on culling it would be quite a bit faster without culling compared to GTX460 - of course, that's not a full GF104 chip so against a 8 SM GF104, clocked a bit higher (I'm still waiting for this...) it would be a bit worse, but still in the same ballpark. If it actually can do 4 tris every 5 clocks in practice that is still more than twice over what we saw with Redwood/Juniper/Cypress. Might not be class-leading but "good enough". Of course, with Barts Pro the numbers would be worse, but not terrible neither.

 

But very unlikely. Why would AMD not implement those in Barts? Just to launch it a month earlier?

 

There's an article at SA, written by LGN, summarising the rumours about the 6800 series. Not much new stuff, but there is following piece of interesting info:

Could that be the reason we see the big difference in PRO and XT price? That means the 6870 should soon drop, unless supply won't meet with demand.

 

Prices, where? I missed that...

 

@ -The_Mask-

In the light of Barts being measured @230mm² and Cayman being rumored to be a 380-400mm² chip, wouldn't you want to adjust your notepad-ideas as far as Cayman's shader count is concerned? What about 2240?

Given your current speculation, a Cayman chip @ the specs you suggested would arguably end up smaller than Cypress ...

 

If Cayman and Barts aren't the same architecture then there's no point comparing their die sizes.

 

I'm just trying to think within the framework of "ideas" -The_Mask- suggested ... it's my layman-way of finding out whether to trust them or not

 

Posted yet? saw this posted somewhere else.

Chinese site with some new pictures and some benches:
http://tech.sina.com.cn/h/2010-10-16/06171529120.shtml

 

Well whaddya know?

Looks a lot like the 6870 will be close enough to the 5870, will probably overclock a lot better, be lower power, 1/3rd cheaper...and everybody will forget all about this renaming shite along with the 5870.

So it's pretty much what happened with GF100 and GF104, give or take a couple of (admittedly important) details.

 

So, can we expect at least 50% over those performance numbers for Cayman?

 

it seems legit

 

if it is correct how did they get a 30% perf increase per mm sq?

 

This looks way more realistic with Barts having 1280SPs rather than measly 960. (but they fu RV840/Juniper specs 144x4 weird) Only bad thing imo is that they could easily add up 2GB parts for 50bucks more even at startup cause RV940/Bart vs. RV870/Cypress have pretty much obsolete-casual gddr5 speced at 1050Mhz only (like old 4890 or worse )
We could saw pretty cheap (hope so) ef-SIX designs based on this chip in three-more like six month period sport 2GB. Even thou they could done that, as wishful thinkers mentioned, right at start it's not like they'd loose some money but i guess this is good enough for them to push envy further into abyss.

And who advocated that 1280SP "new shaders" require 2800mm2? After all they allegedly claim that they're changing logic for the new NI/SI/(we) breed and couldn't that account for memory interface and maybe smaller/squeezed L2 cache also?! (also no sideport and no second CF channel for QuadCFX)

So RV840/Bart will be just as every x30/x40 part (with RV740 being gracious anomaly) just a mainstream gaming card, with maybe just single precision ocl support, while provide great experience for most of casual non-addict market. Just like 4770/4830/5770 were before (x1950pro/x8X0pro in some way), and 3650/4670 being more of an exception based on inherently clogged R600 design and "illogical marketing motto" that new-gen mainstream card *must not* be better than high-end card of previous gen.
And 1280SPs nicely continue ~640SP per unit part that RV740 proofed to be good spot for future designs.

 

Too bad there wasn't fill size of those box pictures.

 

Exactly! I tried to zoom..and zoom...AND zoom some more. To no avail.

 

I think basing this solely on some 3dmark scores is a bit optimistic.
Also, I sort of doubt it overclocks better. I'd suspect it's still designed for around the same clocks, but clock is higher by default already. Plus it seems to achieve the similar performance with a smaller die (hence less transistors) but power isn't actually that much lower from what we saw, hence this suggests it's pushed already pretty hard. Of course, as usual the slower part should overclock better.

 

It's not just that easy, they might look out for naming scheme to enemies lodging ... or if i might axsk should gtx450 be named gts460 cause after all it's just "mainstream version" (gf104) of their infamous big bro gf100 Every gX4 chips in recent past (since g84) carried only "gts" or even measly "gt" branding 8600gts/gt-9600gt---gt240/340 (if we loosely assume gt215 to be gx4 part)

Anyhow, i smell Bart should be willful participant to 28nm dumb shrink test with die size <130mm2 in June next year, so they'll need some spare naming space to offer something in reasonable HD6700 space ... because for the sake of marketing (hopefully) they couldnt sell 130mm2 chip for 151-200USD and retain HD6800 name. Only weird thingy is that they still need to cut off "two non-working pipelines" from fully working 1280SP on RV940 chip, which probably, just as RV770 and RV570 before it, will came out to be healthy die in 90% cases.
Only bad thing in this scenario is that HD6700 will be yet another bw limited 128bit part with maybe 6Gbps gddr5.

Lets hope that nVidia will soon introduce speculating gtx475 cards on fully working gf104 chips, and that amd will be forced to abandon 6850 with laser-cut 160SPs in favor of fully working HD6870

And renaming pricey HD5770 to let say HD6670 along with sizable price cuts to 100USD would be much appreciated especially if they could be easily paired with older 5770 without bios xflash and similar fuss. Renaming it just for the sake of renaming wouldn't be good. And yet better would be that they simply cut down prices and retain original name.

 

Not without cutting bus width.