AMD: R9xx Speculation

[Unknown Soldier]

Kewl. Wonder if the reviewers used the same driver across the board for the 58xx/68xx reviews. Maybe after the new drivers get released?

Haven't seen 10.10 up yet.

*edit* nm

http://www.techpowerup.com/reviews/HIS/Radeon_HD_6850/5.html

NVIDIA: 258.96
ATI: Catalyst 10.7
HD 6800: Catalyst 10.10 Oct 12

 

[no-X]

Darn... Wide-tent gone is indeed rather saddening. Was one of the reasons I would give Radeons extra points in consideration.

MLAA looks fine and dandy, but I don't think anything would match the near-filmic look of WT still.

I agree, it greatly reduced flickering of small objects in motion and crawling artifacts. MLAA cannot replace WT, because artifacting in motion is its weakest point.

I don't understand, why they removed it. Couldn't they just let it be? I hope it will be accessible at least by some kind of 3rd party utility.

 

[Harison]

And what was the result? You would think after 7 "generations" of tessellation hardware they would be blazing trails but instead they're stuck with complaining about nVidia as usual.

Nvidia and their fans are really something Both AMD and Nvidia LOVE tessellation, hell - AMD was the very first to the market (and like 10 years before anyone else if we count first incarnation), the difference comes in benefits vs going overboard (which doesnt really benefit users). Do you see any difference between very good and extreme tessellation in games? My bet is you dont in most cases, if at all.

Example: AA 8x is amazing feature, isnt? On 2560x1600 screen its enough for 99.99% of users, thats the level AMD is promoting with tessellation. Nvidia went overboard with extreme tes., its much like saying AA 24x is THE way to go. Do you see difference between AA 8x and 24x on 2560x1600? My bet is you dont in most cases, if at all.

Thats why its funny to read how AMD tessellation isnt good enough, or how AMD dares to suggest devs to use very good tes. levels (which benefits users), but not extreme (users cant see the diff).

 

[mczak]

And? This is the 5870, too. But what it with Stone Giant? Or the DX11 Samples from SDK?

Anand has numbers for the DX11 Tesselation Sample: http://www.anandtech.com/show/3987/...enewing-competition-in-the-midrange-market/17 - The difference to HD5870 is indeed rather pronounced there.

We look at many different potential configurtions. The two being discussed for Barts were effectively the same (target) cost so they were analysed more closely to establish which way to go.

Interesting that a 50% difference in ROPs is just about equal to a 15% difference in shaders.
I'd prefer a 16 SIMD/32 ROP configuration though (obviously coming at a cost - though I think that would have been pretty much enough for the HD6850/6870 cards to catch up with HD5850/HD5870).

I don't know about the GF104 being obsolete; I was pretty damn surprised at how well the EVGA 460 FTW card performed in Anand's review. It's $10 cheaper than the 6870 at the moment as well.

Looks like very much cherry picked sample to me. Those heavily overclocked GTX460 usually come with quite a bit higher voltage, which brings up their power draw to GTX470 levels - this wasn't the case here.
That said there's no doubt a full GF104 chip could compete quite well with requiring only about a 750Mhz clock - for some reason it's just not happening.

Wow that's a vast difference for such a measly gap.

Yeah, I'm surprised it makes that much of a difference too for just a 15% difference in clock. Wondering how a 6Ghz gddr5 controller is going to look like then...

 

[Jawed]

Interesting that a 50% difference in ROPs is just about equal to a 15% difference in shaders.
I'd prefer a 16 SIMD/32 ROP configuration though (obviously coming at a cost - though I think that would have been pretty much enough for the HD6850/6870 cards to catch up with HD5850/HD5870).

Well, the ROPs have way less bandwidth per cycle than in HD5870, so just increasing the bandwith would probably make HD6870 the same performance as HD5870.

 

[nyr]

Thats why its funny to read how AMD tessellation isnt good enough, or how AMD dares to suggest devs to use very good tes. levels (which benefits users), but not extreme (users cant see the diff).

AMD clearly don't think it's good enough. Otherwise why would Cayman be featuring Gen 8 tessellation with improved scalability?

 

[Harison]

I'm very excited about MLAA. Almost no drop in performance, should work on all new and old games. No wonder AMD is pushing it to the market, respect for the positive endeavors. :smile:

 

[Dave Baumann]

Well, the ROPs have way less bandwidth per cycle than in HD5870, so just increasing the bandwith would probably make HD6870 the same performance as HD5870.

The ROPS also have more idle time on Barts.

 

[CarstenS]

You are not taking it into account in the settings that you are comparing.

Rest assured that we do. Please note that we're not comparing settings purely looking at the sliders' designation but try to analyse the output we're seeing on screen.


And also note that we indeed honour your commitment to give control back into the hands of the users. I even dedicated my personal column in the mag to it. Really - IMHLO it really outweighs the changing of default settings for Cypress-Users.

 

[Dave Baumann]

AMD clearly don't think it's good enough. Otherwise why would Cayman be featuring Gen 8 tessellation with improved scalability?

Barts is designed to address a specific market segment, which is not necessarily the case for other products.

 

[Harison]

AMD clearly don't think it's good enough. Otherwise why would Cayman be featuring Gen 8 tessellation with improved scalability?

Tech. is always improving, like my example with AA 8x on 2560x1600, even though its enough, we get higher AA levels and now MLAA as well.

Or look from another perspective, if AMD/Nvidia wouldnt be improving their cards, what they would sell to you? 5800 gen. was more than enough for vast majority of users, yet we get 6900 to replace it, next year 7900, etc. Business as usual.

 

[Bouncing Zabaglione Bros.]

AMD clearly don't think it's good enough. Otherwise why would Cayman be featuring Gen 8 tessellation with improved scalability?

Your suggestion is that every lower product isn't "good enough" because there are better products that costs more higher up the range?

 

[Unknown Soldier]

AMD clearly don't think it's good enough. Otherwise why would Cayman be featuring Gen 8 tessellation with improved scalability?

Silly child, think of the 68xx series as if it was 67xx and all your problems will go away.

 

[no-X]

AMD clearly don't think it's good enough. Otherwise why would Cayman be featuring Gen 8 tessellation with improved scalability?

Do you think, that R9700's anti-aliasing implementation wasn't good enough (esp. when compared to competition), hence ATi decided to improve it in the R9800?

 

[{Sniping}Waste]

Dave, y'all did a great job with Barts. The performance for the specs is a nice improvement over the HD5XXX and even did it with less power needed. The only problem most are seeing is the names of HD6850/HD6870. I see why y'all didn't use HD6750/HD6770 because the performance is over doubled. I think the better names would have been Barts pro = HD6830 and Barts XT = HD6850. A cut down Cayman could take the HD6870 name.

 

[nyr]

Barts is designed to address a specific market segment, which is not necessarily the case for other products.

Indeed, but in the 5000 series the common tessellation over the entire range was presented as a virtue, while nvidia's scaling was presented as being stupid. Although this may have been CharlieD's doing rather than an offical AMD position.

But my issue mainly comes more from the recent statements from Richard Huddy that there is a minimum triangle size beyond which there is no point in tessellating. If it's not worth tessellating past that point now, why will it suddenly become worth doing it next month?

 

[Broken Hope]

There is no "Cat AI standard" going forward. The optimization behaviour has been unified across all boards from Cat 10.10 onwards, so the default beviour on Cypress is different, but you can get back to Cat 10.9 quality on Cyress by changing the slider to high quality.

Am I missing something, I just installed the Catalyst 10.10 drivers that are currently uploaded to the AMD servers and I don't have multiple quality options for Catalyst AI on my 5870, just standard and advanced like normal.

[edit]

Also what have you guys done with the colours? I have a horrible pink tint on what was a perfectly calibrated display using a hardware colorimeter.

 

[Kaarlisk]

Well, the ROPs have way less bandwidth per cycle than in HD5870, so just increasing the bandwith would probably make HD6870 the same performance as HD5870.

That's memory bandwidth, if I understand correctly?

Also, has there been a review where the impact of having two "Ultra-Threaded Dispatch Processors" is discussed/explained?

 

[Kaarlisk]

Dave, y'all did a great job with Barts. I see why y'all didn't use HD6750/HD6770 because the performance is over doubled. I think the better names would have been Barts pro = HD6830 and Barts XT = HD6850. A cut down Cayman could take the HD6870 name.

That would have made HD6870 either too expensive to make, being based on a much larger chip with a more expensive board, or a salvage part with low quantities on sale.

Instead, AMD has succesfully "rebooted" their naming/midrange, and they now have very performant and competitively priced parts in the $180-240 range.

 

[Man from Atlantis]