AMD: R9xx Speculation

Synthetics would have to be really representative of real world in this case...

Remember a 1920SP/850MHz is worth twice GTX580's peak throughput and Cypress previously already posted some impressive numbers in synthetic tests such as 3DMark PS test.

I don't know if we'll see that huge of a gap, but it's further ahead of Cypress than GTX580 of 480 (20% shaders, VLIW4 instead of VLIW5, dual dispatcher/tessellator and reworked back-end vs 1 more SM and 10% higher frequency...), so GTX580 performance is a very low estimate.

 

1. You forgot to factor in clocks, HD 6870 is at 900 Mhz vs 5870 at 850. Clock for clock, HD 6870 is ~10-15% slower.

2. You could also look at it from another viewpoint, HD 5870 with 40% more shaders is only 10% faster than 6870. So clearly shader power isnt everything

3. The rumours say ROP's will remain at 32 along with 256 bit memory interface. Could be a possible bottleneck

4. All these simple shader extrapolations are screwed up anyway because of the change to VLIW4(apparently performance is similar, but do we know how much, ie 5,10,15%? and whether its a performance gain or loss)

 

This is confusing, they bring out a new Arch with the same naming as the current Gen?

Or Do they plan to produce a 7800 by renaming the current 6900. And make lower end of 6900 series as 7xxx series?

 

As a simple rule of thumb you can use ~10%. That number works in both directions. The same number of ALUs will deliver ~10% more effective shader power on average. And the same number of VLIW4 units will deliver ~10% less performance than that number of VLIW5 units on average. And all of that in 10% less die area estate (shader area only of course) for about the same performance according to AMD

Edit:
The whole thing is of course workload dependent. But it is a somewhat conservative guesstimate for the average.

 

Depends on what the target is. I think the launch windows is Nov - Jan, for silly season sales, with the biggest part of the bump right now through to Christmas.

In a perfect world, where my opinion mattered, I'd want the mainstream performance parts ready for back to school (late aug), and high end parts hitting just before thanksgiving, ultra just before Christmas/Qwanza/Hannukah/Winter Solstice. Desktop and media consumption/casual parts intro in Jan/Feb for new OEM and Prebuilts, then a mild refresh/update to key mainstream & high end parts around april/may (tax refund, end of school dads/grads summer fun). Hopefuly this hits the consumer spending peaks and people who want to upgrade/update/replace can do so.

So, this Q we've seen NVIDIA hit Nov with a new high end part, and rumors of another one right around the corner. AMD have a mainstream performance part in market, later than 'back to school' but in the holiday sales season. We've got Cayman rumored for this month, with AMD putting it on their road map for Q4, and Antilles due Q1, possibly CES intro?

I'm not willing to call AMD late yet, although Antilles is late to market unless it generates is own sales bump from being filled with awesome win and kittehs, but nowhere near as late as GF100 was for the Windows 7 / 09 holiday bump, although it still managed to hit a bump.

As for performance, historically I think a precedent of +30-40% year on year for the high end is neither too pessimistic or optimistic, but I hope to be surprised.

 

Well I've moaned before that the HD68xx is hardly worth the HD6xxx (or NI) name.
But certainly retail names are not really based on internal architecture, they might be based on features (e.g. you get UVD3), plus some purely marketing reasons (to have complete new lineup, rename some cards where you don't have any new chips). Can't say I like it, but it's not the end of the world.

 

Mansell and Hamilton

 

With Antilles becoming Senna or Fangio or Schumacher? Or will they stick to British drivers?

If anyone wonders what Jawed referred to:


coming from here.

Edit:
Judging from a comparison of the placements of the "x32" for the memory chips of the HD6970/50 and the fact they used a proportional font (variable width) I speculate the HD6950 may also feature 2 GB RAM. Or is a "32" wider than "64" to compensate for the slimmer "1" compared to "2"?

Actually I expected the HD6950 to have just 1 GB RAM.

 

In general, typefaces have all the numerals occupy the same space, so 1 and 2 wouldn't show a difference:

11111111
22222222

64646464
32323232

 

I didn't call 'em late, I said "delayed". I actually agree with your assessment/analysis, but the bump back delay has added something to this release in the way of expectations.

I do to, I'm still a cockeyed optimist. (cue PJ or LB... )

 

Hmm Mansell faster than Hamilton?

 

Schumacher is reserved for Antilles then.

 

August screenshot seems like it was real

It's worth noting that this article's screenshot:

http://vr-zone.com/articles/alleged-ati-radeon-hd-6870-3dmark-vantage-benchmark-leaked/9692.html



Has a device ID that matches the documents above, i.e. Cayman XT.

So it seems the chip has been floating around since August.

The two graphics tests are slower than GTX580:

http://www.legitreviews.com/article/1461/18/

40 and 35 fps respectively.

This August screenshot did have some credibility at the time, so it seems in retrospect to be quite real.

 

You are right, most fonts are constructed this way. Never came across that fact

 

I think the list with the device IDs appeared just a few days before that alleged screenshot.

 

That Panda looks like it's being... pleasured.

 

And the tested CPU... too...

 

Why the description above the specs regarding Eyefinity is worded toward Cypress/5800 series?

 

Copypaste or another fake

 

did anyone mention the 15th?