AMD's Radeon HD 6870 benchmarked? (updated more screens)

[Xoulz]

pff come on, you know thats not how it works :P
especially in a tesselation geometry benchmarks pixels dont matter that much...

Are you saying the resolution doesn't matter, when dealing with tessilation..?

[Calmatory]

Are you saying the resolution doesn't matter, when dealing with tessilation..?

At least I am. Resolution (UNLESS IT CHANGES FIELD OF VIEW!) doesn't have ANYTHING to do with tesselation. The units which handle tesselation do not handle per-pixel operations, as far as I know, so there is no extra work for those units. The work remains the same regardless of the resolution.

However, obviously even with tesselation, the resolution has impact on the framerate, because there is a limit how many operations the ROPs can do, and how many operations the TMUs can do. With bigger resolution they need to do more. And they also need a bigger share of the memory bandwidth, too.

[Lanek]

At least I am. Resolution (UNLESS IT CHANGES FIELD OF VIEW!) doesn't have ANYTHING to do with tesselation. The units which handle tesselation do not handle per-pixel operations, as far as I know, so there is no extra work for those units. The work remains the same regardless of the resolution.

However, obviously even with tesselation, the resolution has impact on the framerate, because there is a limit how many operations the ROPs can do, and how many operations the TMUs can do. With bigger resolution they need to do more. And they also need a bigger share of the memory bandwidth, too.

offcourse it will, tesselation impact on the anti-aliasing and anistrope filtering....... indirect lightning, and reflect............... It was not the case when ATI have release the API cause driver was optimised for it, but as Microsoft have choose a diferent path in DX11............. it's impact it a lot.......... For make it simple, tesselation don't kill the framerate, but the addtion of the filters uppon the surface tesselated do it.......

[Xoulz]

At least I am. Resolution (UNLESS IT CHANGES FIELD OF VIEW!) doesn't have ANYTHING to do with tesselation. The units which handle tesselation do not handle per-pixel operations, as far as I know, so there is no extra work for those units. The work remains the same regardless of the resolution.

However, obviously even with tesselation, the resolution has impact on the framerate, because there is a limit how many operations the ROPs can do, and how many operations the TMUs can do. With bigger resolution they need to do more. And they also need a bigger share of the memory bandwidth, too.

Per-pixel..? Bro, resolution is ratio. More is more... 1920 x 1200 is more than 1920 x 1080. Any benchmark will change it's fov to render @ said resolution, otherwise it would just stretch scenes & be dysfunctional as a benchmark.

More to tessellate, more work to be done. Aren't you dismissing the obvious?

[Chumbucket843]

At least I am. Resolution (UNLESS IT CHANGES FIELD OF VIEW!) doesn't have ANYTHING to do with tesselation. The units which handle tesselation do not handle per-pixel operations, as far as I know, so there is no extra work for those units. The work remains the same regardless of the resolution.

are you familiar with a z-buffer? triangles directly affect rasterization and z-buffer performance. because of the way z-buffer algorithms are optimized(hierarchy, with quadtrees, coarse grained culling) tiny triangles hurt efficiency which tessellation creates.

i am saying that the # of triangles and their properties are directly related to rendering speed.

[saaya]

are you familiar with a z-buffer? triangles directly affect rasterization and z-buffer performance. because of the way z-buffer algorithms are optimized(hierarchy, with quadtrees, coarse grained culling) tiny triangles hurt efficiency which tessellation creates.

i am saying that the # of triangles and their properties are directly related to rendering speed.

why do you guys continue to discuss semantics though?
the original point was that one score was 1920x1080 and the other 1920x1200 and that they were not comparable... to which i and others said, the difference in unigine between those two resolutions should be tiny and the scores would be somewhat comparable...

and they are... somebody compared 1080 vs 1200 and the performance is almost the same...

[saaya]

Are you saying the resolution doesn't matter, when dealing with tessilation..?

im sure it does, but from what ive seen the difference is very very low...

especially when you go from 1920x1080 to 1920x1200, which is what the original point was...

Well, one thing saaya.

tesselation performance is not directly correlated with resolution when doing a unigine test, but.

Rendering is done per pixel. More pixels = more area to render, which means more shading/displacement/geometry etc....

I can tell you about normal CGI rendering, which shares a lot of similarities.

Displacement takes quite a lot of time to render and when the resolution is smaller, rendering goes faster. If you leave the same displacement and up the res, you increase linearly the rendering time, because the objects being displayed/rendered share a bigger pixel space. Displacement is done on screen, not for the whole scene, it would be crazy to do it like that.

So, resolution increase means also lower performance, even in Unigine.

thx for the details
but im sure you will agree that the difference between 1080 and 1200 is tiny

Fake

so its really fake then?

I don´t know if all the leaked numbers are fake, but real performance must be very good. Otherwise i don´t think that AMD would kill the Ati brand.

good point... but as long as cayman outperforms gf104 they own the highend with a dual cayman card, so... they dont necessarily have to kick 4ss to feel confident about the 6000 series... hmmm

[spursindonesia]

.............

so its really fake then?

good point... but as long as cayman outperforms gf104 they own the highend with a dual cayman card, so... they dont necessarily have to kick 4ss to feel confident about the 6000 series... hmmm

Saaya, if the new chips are not really fast enough over the current line ups, so like you've previously suggested, then why they have to bother R&D & creating this whole new family ? DX 11 strengthening can wait until 28 nm arrives, can't they ? We know in current games, Evergreen is more than good enough on average.

Evergreen while long in the tooth is holding the fort just fine, and since Evergreen fastest chip is actually smaller than nVidia's current best perf/watt/die size area performer, GF 104, should the supply constraint abated, ATi could easily engage a price war for market share expansion, if needed.

These new chips IMHO have to be considerably faster while adding efficiency in the process, i know, that's quite a feat should ATi successfully pull it all the way out, but i think they have the capability & room to be so. Then, a new generation tagline would be well deserved & justified.

Regards.

[tajoh111]

Saaya, if the new chips are not really fast enough over the current line ups, so like you've previously suggested, then why they have to bother R&D & creating this whole new family ? DX 11 strengthening can wait until 28 nm arrives, can't they ? We know in current games, Evergreen is more than good enough on average.

Evergreen while long in the tooth is holding the fort just fine, and since Evergreen fastest chip is actually smaller than nVidia's current best perf/watt/die size area performer, GF 104, should the supply constraint abated, ATi could easily engage a price war for market share expansion, if needed.

These new chips IMHO have to be considerably faster while adding efficiency in the process, i know, that's quite a feat should ATi successfully pull it all the way out, but i think they have the capability & room to be so. Then, a new generation tagline would be well deserved & justified.

Regards.

Profit margins. AMD with the current price makes quite a bit of money but what if they could charge about a hundred dollars more across the line. They would be making a crap load more, probably a hundred million more with the number of cards they already sold.

With the rates that these cards are selling out, AMD might be thinking they released the 5xxx series at too low of a price and they can not outright raise the price themselves too much(the most they could get away with was the 20 dollar raise they did).

What they could do is release a somewhat modified card which has better performance that atleast matches the competition and somewhat beat's it and increase the price of the line 100 dollars(and these cards likely cost the same to produce considering improved yields). Not taking huge design risk allows them to develop the card way faster which is probably the most important thing considering the time between now and 28nm is when NV will have nothing to respond with in the high end market beside the gtx 490.

By keeping this performance in check and not doing something crazy to this generation: it allows them to make a new product for cheap(low R and D), keep current 5xxx customers happy by not feeling that they were duped into buying a slow early 5xxx and it also prevents these owners from flooding the market with used 5xxx cards which might prevent a new sale of a 6xxx series.

According to steam, close to half the 5xxx cards are the 58xx variants. Thats 8 million cards considering 16 million cards have sold. Both the 5850 and 5870 have prices of 259 and 379 MSRP but have sold for atleast at the beginning at $350 and $450. What if AMD simply refreshed the line with about 20% more performance, released at better quantities so that the card sold at MSRP and they reaped the benefits rather than the retailers who gouged us.

This is what AMD should have done in the first place and I think they realized this.

This quarter, the AMD graphic division made 40 million dollars net which is pathetic considering the extreme monopoly they had. NV had quarters where they made hundreds of millions of dollar during their prime. Just imagine how much difference to this net profit if they charged 100 dollars more, if AMD for example charged 100 dollars more and sold 10 million cards not including the lowend, depending on the deal they had with partners, your looking at atleast 200 million more in your net profits even with R and D, money shared with partners.

Fermi has been such a PR disaster for the consumer because of heat and energy concerns that the rate of the 5870 and 5850 don't seem to be slowing down that much and are still selling out.

They could release a card that just as fast as the gtx 480 add 100 dollars to the 58xx current price and it would still sell if they avoided the energy and heat output of the gtx 480.

They have alot of momentum so people might just buy an AMD card just because it is new and they haven't jumped on the directx 11 fence this generation.

Considering SI was originally design for 32nm and considering how quickly they got this line up together I am thinking this card just from a time perspective has to be way way more cypress than NI, which lead me to believe they might improve performance 20-30 percent but nothing crazy like the fake tesselation score in unigine.

[LightSpeed]

Snip

Tajoh, ive been hearing that Cayman was originally designed for 28/32 nm as well.

Thinking about it some more, it makes sense. As GPU's take years to go from design to being finalised, I am sure the architecture after Cypress was designed with 28/32 nm in mind. Its not like the transition from 45 to 32 nm was taking too long so AMD just decided to modify the Cypress core here and there and sell it as Cayman in a few months. Also, the fact that 67xx seem to be modified Cypress cores means that Cypress would be cheaper to manufacture than a dumbed down Cayman cores, which might happen due to the fact that Cayman will be bigger in size compared to Cypress since, again, it was designed for 32nm.

Power consumption apparently seems to have gone up in comparison to Cypress as well. Hmm.

Also, you do need to realise that 67xx are going to replace the 58xx cards. I doubt AMD would price their midrange 67xx cards at current 58xx prices. Hence from what I feel, 68xx will be priced slightly more than 58xx, and 67xx are going to fill up the $150 - $200 market.

Of course I might be wrong

[tajoh111]

Tajoh, ive been hearing that Cayman was originally designed for 28/32 nm as well.

Thinking about it some more, it makes sense. As GPU's take years to go from design to being finalised, I am sure the architecture after Cypress was designed with 28/32 nm in mind. Its not like the transition from 45 to 32 nm was taking too long so AMD just decided to modify the Cypress core here and there and sell it as Cayman in a few months. Also, the fact that 67xx seem to be modified Cypress cores means that Cypress would be cheaper to manufacture than a dumbed down Cayman cores, which might happen due to the fact that Cayman will be bigger in size compared to Cypress since, again, it was designed for 32nm.

Power consumption apparently seems to have gone up in comparison to Cypress as well. Hmm.

Also, you do need to realise that 67xx are going to replace the 58xx cards. I doubt AMD would price their midrange 67xx cards at current 58xx prices. Hence from what I feel, 68xx will be priced slightly more than 58xx, and 67xx are going to fill up the $150 - $200 market.

Of course I might be wrong

I could see if AMD doesn't update the 5770 them still keeping the 6770 range at the low price they are at. However, if it performs 25% better, I could see the 6770 being knocked up 50 dollars. If the 6770 performs like a 5830 or a gtx 460, it can totally get away with a price tag of 179.99 and the 6750 being 150. AMD can get away with such pricing because they have the momentum right now and Nvidia's low range products are looking mediocre.

If the 5xxx series is off the market, the consumer will be basically forced to buy AMD's line if they want something new at all or buy NVidia's slightly cheaper but worse performing line with higher power and heat.

The later they don't mind because they allow them to be reestablished as the premium brand in addition it allows their competitor to make some money just not that much.

Making money is the number one priority of a company, and getting something as quick to the market ASAP when the competition is vulnerable is the most important thing during that time. Nvidia is going to jump on 28nm ASAP when it comes out and AMD going to need a line at 28nm at that time too. This is likely going to be late 2nd quarter or early in quarter 3 of next year.

I cannot see AMD spending a full R and D budget and more to rush this product on a product that's going to only have a 9 month life cycle(which will be likely the shortest lifecycle ever for a lineup). They really don't need to blitz NV at this point. They are selling like crazy as is, and by keeping performance in check, it gives more incentive for people to upgrade when the 7xxx generation happens. Considering that this chip has to be bigger than the chips as is, to perform better, its not the place to take risks as we have seen with the TSMC 40nm. Its better to be safe so you can have predictable power outcomes.

I could see Nvidia still selling during this period with a 384 shader part at 249, gtx 480 at $399 and a gtx 490 at a price of 499.

Too much performance really is going to lead to gouging which really doesn't help AMD as a brand to profit as seen from their last quarterly statement. At this point timing is more important. The longer AMD can get away with big margins, the more money they will make.

[Chumbucket843]

Making money is the number one priority of a company, and getting something as quick to the market ASAP when the competition is vulnerable is the most important thing during that time. Nvidia is going to jump on 28nm ASAP when it comes out and AMD going to need a line at 28nm at that time too. This is likely going to be late 2nd quarter or early in quarter 3 of next year.

that's almost never done and for good reasons. also a gf100 based on 28nm could be easily 40% faster with relatively little effort.

I cannot see AMD spending a full R and D budget and more to rush this product on a product that's going to only have a 9 month life cycle(which will be likely the shortest lifecycle ever for a lineup). They really don't need to blitz NV at this point. They are selling like crazy as is, and by keeping performance in check, it gives more incentive for people to upgrade when the 7xxx generation happens. Considering that this chip has to be bigger than the chips as is, to perform better, its not the place to take risks as we have seen with the TSMC 40nm. Its better to be safe so you can have predictable power outcomes.

R&D is a fixed expense. it doesnt change much. but really it's not a quick paced game like you think. development of gpu's is 3-4 years so they arent going to base decisions on how the competition is performing 3 years from now, they are just going to reach the goals of the chip. i.e. performance, area, cost, etc.

[spursindonesia]

Profit margins. AMD with the current price makes quite a bit of money but what if they could charge about a hundred dollars more across the line. They would be making a crap load more, probably a hundred million more with the number of cards they already sold.

With the rates that these cards are selling out, AMD might be thinking they released the 5xxx series at too low of a price and they can not outright raise the price themselves too much(the most they could get away with was the 20 dollar raise they did).

What they could do is release a somewhat modified card which has better performance that atleast matches the competition and somewhat beat's it and increase the price of the line 100 dollars(and these cards likely cost the same to produce considering improved yields). Not taking huge design risk allows them to develop the card way faster which is probably the most important thing considering the time between now and 28nm is when NV will have nothing to respond with in the high end market beside the gtx 490.

By keeping this performance in check and not doing something crazy to this generation: it allows them to make a new product for cheap(low R and D), keep current 5xxx customers happy by not feeling that they were duped into buying a slow early 5xxx and it also prevents these owners from flooding the market with used 5xxx cards which might prevent a new sale of a 6xxx series.

According to steam, close to half the 5xxx cards are the 58xx variants. Thats 8 million cards considering 16 million cards have sold. Both the 5850 and 5870 have prices of 259 and 379 MSRP but have sold for atleast at the beginning at $350 and $450. What if AMD simply refreshed the line with about 20% more performance, released at better quantities so that the card sold at MSRP and they reaped the benefits rather than the retailers who gouged us.

This is what AMD should have done in the first place and I think they realized this.

This quarter, the AMD graphic division made 40 million dollars net which is pathetic considering the extreme monopoly they had. NV had quarters where they made hundreds of millions of dollar during their prime. Just imagine how much difference to this net profit if they charged 100 dollars more, if AMD for example charged 100 dollars more and sold 10 million cards not including the lowend, depending on the deal they had with partners, your looking at atleast 200 million more in your net profits even with R and D, money shared with partners.

Fermi has been such a PR disaster for the consumer because of heat and energy concerns that the rate of the 5870 and 5850 don't seem to be slowing down that much and are still selling out.

They could release a card that just as fast as the gtx 480 add 100 dollars to the 58xx current price and it would still sell if they avoided the energy and heat output of the gtx 480.

They have alot of momentum so people might just buy an AMD card just because it is new and they haven't jumped on the directx 11 fence this generation.

Considering SI was originally design for 32nm and considering how quickly they got this line up together I am thinking this card just from a time perspective has to be way way more cypress than NI, which lead me to believe they might improve performance 20-30 percent but nothing crazy like the fake tesselation score in unigine.

Thx for the lengthy response, appreciated the thought & gesture.

Regarding your opinion, it has merit of its own, but i dunno man, i don't think ATi will make a whole family refresh just to keep profit margin by creating a bigger & more power demanding chip at the top and just relegating the old chips downward + renaming them nVidia's style. Remember RV 790 ? A new SKU with relative minimal R&D to matchup better with competitor's better SKU, it still kept the same family nomenclature. Only time will tell the truth (will be unveiled in the near future i presume) & speculation is fun, like what Saaya said earlier.

About the leak numbers, i think those are quite possible, but for me personally, the Vantage & Crysis score seems more believable & more probable. Why ? Because i do have a strong feeling regarding the added efficiency of the rumored, new 4D SP array design, applied on Cayman with 1920 SP (480*4). Regarding tesselation, while the smart money will put on a stand that ATi should have improved its current mArch relative dismal geometry & tesselation capability, i really don't have a clue on how they could achieve that.