Radeon HD 7970 Tessellation Performance Figures Surface

[Benetanegia]

We also dont know if the CUs are locked to clock speed or they went with a more NVIDIA approach and sped them up.

They are still locked if we are to believe the rumors.

Again I don't expect shaders themselves to be very different from what they were.

So there is a slither of hope, very slim though.

Yeah, but I wouldn't count on that. I'd rather be surprised than dissapointed. Again.

And the complete lack of any performance slides other than tesselation performance is not good really.

 

[air_ii]

With 2048 SPs you really need them SPs to be a lot faster/efficient than they were if you want the card to be significantly faster. It is fairly posible to have much faster SPs, but like with BD we again depend on a net "IPC" increase in order to get a significant performance gain. Since I don't want to believe in fairies again, I'm counting on SPs being fairly equal hence with 30% more SPs and 30% more TMUs and similar clocks, 30% performance increase is kinda the maximum I would expect. According to OBR numbers, net "IPC" or perf-per-shader might actually be slightly down and wihle it was not something I would have expected, it's not really something completely imposible.

Assuming the same speed of the CUs, you should at least factor in the inefficiency (ie underutilisation) of the D5/D4 VLIW architecture, which should not be an issue with GCN.

This time around OBR does seem to have access to slides, which look legit. Who knows, maybe he does have a source. But then again, the last time AMD put a launch forward was with the 4870/50, which turned out to be pretty nice. So if the rumours are correct, we'll know in a week's time.

 

[Benetanegia]

Assuming the same speed of the CUs, you should at least factor in the inefficiency (ie underutilisation) of the D5/D4 VLIW architecture, which should not be an issue with GCN.

That's the point, for graphics there is no big inefficiency. There was some for VLIW5 (~80% efficiency) but for VLIW4 it should be much better. I've not seen numbers for VLIW4 but for VLIW5 the average utilisation was 3.9 out of 5. Of course for VLIW 4 it won't be the same, it has to be slightly lower because 3.9 was achieved thanks to sometimes doing 5 of 5, but at least 3.6 is probable, also accounting for the performance compared to Barts and Cypress, etc.

You have to realise (everybody needs to) that AMD (and Nvidia a long time ago) has abandoned VLIW, not because it's inefficient for graphics, but because it's inefficient for GPGPU. From a graphics standpoint, they are not doing this because they want, they are doing it because they need to, because that's the only option if they want to ever enter the HPC market.

 

[MxPhenom 216]

I am not sure how well nVidia did in Tessellations but normally only the stuff that AMD want to get leaked actually gets leaked, from this i would read that the Tess improvement is 7970's biggest improvement.

I hope i am wrong.

Im confused. why would the massive tesselation improvement with the 7970 be a bad thing. Tesselation/Geometry processing has always been AMD card's weakest link.

 

[Rowsol]

Well, heaven is a great benchmark and that performance increase just made me hard.

 

[Benetanegia]

Im confused. why would the massive tesselation improvement with the 7970 be a bad thing. Tesselation/Geometry processing has always been AMD card's weakest link.

He is refering to the fact that they are only talking about tesselation improvement. With less than a week until release, they have not leaked a single performance chart other than that.

 

[15th Warlock]

To be honest with you, I'm very excited about these cards, and have really high hopes for them, I want a new DX11 card for my HTPC, and the XFX 6970 I recently ordered was DOA, so if these cards are half as fast as they are being hyped to be, I would gladly get one to replace my 285s, if not, we know AMD will price them accordingly and that'll lower the prices of cards like the 570 and the 6970, and I may give it another shot , so either way, it's a win-win scenario.

AMD needs a winner here, and although I wasn't happy about them buying Ati a few years back, you have to admit their latest GPUs and APUs are the only thing keeping both Intel and Nvidia from obscenelly increasing their prices, and that can only be good for us as costumers

 

[MxPhenom 216]

He is refering to the fact that they are only talking about tesselation improvement. With less than a week until release, they have not leaked a single performance chart other than that.

Probably because with the new architecture of these cards thats the biggest improvement. I mean off course there is improvement in the other areas and a big one im pretty sure, but Tesselation has never been AMD cards thing. So I would think that they would try to improve that a bit for the new cards.

 

[PopcornMachine]

I'm a fan of ATI/AMD graphics cards, so I hope these figures are accurate. But AMD earned my complete distrust of any of their marketing ploys with the Bulldozer fiasco.

Perhaps things are better now that they replaced their whole marketing team, but perhaps not.

Fortunately I don't need a new graphics card. In any case, I would advise anyone to wait an see independent reviews before purchasing anything from them.

 

[Casecutter]

Here’s the problem it could be disinformation plant by whomever, ATI never comments on the validly of such leaks… will not confirm or deny
All this would only indicate (if not tampered) is AMD has enough faith that tessellation will no longer be trailing Nvidia, and appears they have isolated and repaired that bottle neck. I look at it as a positive and one less item to concern ourselves with.

 

[air_ii]

That's the point, for graphics there is no big inefficiency. There was some for VLIW5 (~80% efficiency) but for VLIW4 it should be much better. I've not seen numbers for VLIW4 but for VLIW5 the average utilisation was 3.9 out of 5. Of course for VLIW 4 it won't be the same, it has to be slightly lower because 3.9 was achieved thanks to sometimes doing 5 of 5, but at least 3.6 is probable, also accounting for the performance compared to Barts and Cypress, etc.

You have to realise (everybody needs to) that AMD (and Nvidia a long time ago) has abandoned VLIW, not because it's inefficient for graphics, but because it's inefficient for GPGPU. From a graphics standpoint, they are not doing this because they want, they are doing it because they need to, because that's the only option if they want to ever enter the HPC market.

VLIW was good for pushing vectors and points around. But modern engines have evolved way beyond that. There's lots of other calculations (shading, occlusion, etc) that VLIW architectures are not best suited for. So yes, while I realise that VLIW is good for geometry processing, I still believe that other workloads seen in games tend to be more GPGPU.

 

[Steevo]

ATI's cards have always been competitive, unlike some AMD CPU's, and I'm sure they are releasing this to gain any additional foothold they can with a competitive product offering.


Their software/support is/has always been their downfall, not the drivers, but the BS they promise and never deliver on, or hey, yeah, $600 for that card, and now buy this and this and this to make good on our advertisements.



Fortunately OpenCL and a clean "C" path for programmers will make everything better as long as MS buys in, and with it becoming part of the new DX on 8 it will be a shoo in for AMD and Nvidia.

 

[Benetanegia]

Probably because with the new architecture of these cards thats the biggest improvement.

That's the point. Of the games/benches shown there, which ones are getting the biggest improvement? Heaven and Crysis 2. The ones which are more heavily tesselated. Less tesselation, less performance improvement. If Heaven is on the Extreme tesselation preset, those results are not much better than a GTX 580 either.

The rest use less tesselation and their performance is shown as only 50% over HD6970. With no tesselation at all performance is probably significantly lower than 50% and there's the fact that this are the way way too much overinflated AMD PR slides, as usual. You have to take 20% off of those numbers to get the real thing.

Remember this?

VLIW was good for pushing vectors and points around. But modern engines have evolved way beyond that. There's lots of other calculations (shading, occlusion, etc) that VLIW architectures are not best suited for. So yes, while I realise that VLIW is good for geometry processing, I still believe that other workloads seen in games tend to be more GPGPU.

95% of games are console ports and that's not going to change. VLIW is perfectly well suited for games. You might see a 5-10% efficiency gain going with 1d shaders, you may also loose a 10-20% simply because it's a new architecture also.

But, well, I don't know why people say AMD's marketing department sucks. 1 year ago going with this exact same architecture approach was a really really bad idea, it added an unsourmountable ammount of transistors for no gain. Today, not only does not hurt performance, it will magically make it a lot better! AMD's marketing department sucks? They are gods.

 

[TheoneandonlyMrK]

Tesselation/Geometry processing has always been AMD card's weakest link.

ATI were doing tesselation way ,way before nvidia and are on generation 5-6 or something nvidia were lagging behind and simply added i think 8 -16 tesselation engines into fermi to catch up, which it did but amd havent gone the same route and just thrown more tess units in , cos theirs are just better

and its slightly unimportant anyway because no ones using tesselation right, as it is check crysis 2,s underfloor oceans etc 20-30% increase from AMD would be fine if they get 50%-4x in some games hows that bad

and why do some fags spend their eves shitting on others dreams for a buzz wagwan wi that fanboys do my nut

and imho the move to compute units and MIMD instead of simd and vliw is due to the constant switching between compute and rendering tasks taking a toll on performance ,another aspect of 7xxx performance increase will be multi threaded capabillty ie it can run more then 1 thread at the same time and also its supprort of iommu may up performance too all in its worth keeping an eye on

 

[Casecutter]

marketing ploys with the Bulldozer fiasco

Again, “ATI” never comments on the validly of such leaks… will not confirm or deny.

But the management AMD (CPU side) that was there a year ago had "loose lips" and we known what results. I think Rory and a new PR structure will hold to what ATI side has done, which more often has worked for them, strategic plants, orchestrated talking points, controlling the presentation and effect.

That was the issue BD… A Peacock parade (executives) flapping their wings way too early and not on any talking points. Then the bad news and egos and finger pointing… then everything went out of control. Had AMD came in and calmed waters, clarified the massage, tempered expectations and not say anything (or not used the FX designation), upon release even when late it could have gone better. Rory was right to shake down the PR side… for a part of that disaster.

 

[MxPhenom 216]

ATI were doing tesselation way ,way before nvidia and are on generation 5-6 or something nvidia were lagging behind and simply added i think 8 -16 tesselation engines into fermi to catch up, which it did but amd havent gone the same route and just thrown more tess units in , cos theirs are just better

and its slightly unimportant anyway because no ones using tesselation right, as it is check crysis 2,s underfloor oceans etc 20-30% increase from AMD would be fine if they get 50%-4x in some games hows that bad

and why do some fags spend their eves shitting on others dreams for a buzz wagwan wi that fanboys do my nut

and imho the move to compute units and MIMD instead of simd and vliw is due to the constant switching between compute and rendering tasks taking a toll on performance ,another aspect of 7xxx performance increase will be multi threaded capabillty ie it can run more then 1 thread at the same time and also its supprort of iommu may up performance too all in its worth keeping an eye on

I guess, but Ati cards ever since the HD2900 have been running virtually the same architecture. jsut added new ram controllers, more Sps and tweaked it in the 6xxx series for one more tesselation engine. From the Gt2xx to the GT4xx(even though the 4xx series is a giant chip to suck power, not nessecarily perform) was new and now the geometry processing of Fermi still is not beat by AMD offering.

 

[erocker]

Tesselation is overrated and overused/not implemented correctly by some.

 

[pantherx12]

Tesselation is overrated and overused/not implemented correctly by some.

I don't think tessellation is over rated, definitely not used in the right ways though.

IMO tessellation should be used when your right up close to something, mostly on things like bricks/details and things like that.

No need for it to smooth edges and stuff as that's not all that noticeable.

Also dynamic tessellation so you can smash a guys face in would be nice.

 

[erocker]

I don't think tessellation is over rated, definitely not used in the right ways though.

IMO tessellation should be used when your right up close to something, mostly on things like bricks/details and things like that.

No need for it to smooth edges and stuff as that's not all that noticeable.

Also dynamic tessellation so you can smash a guys face in would be nice.

I agree with your post, I'm saying it's currently overrated in it's current form and/or the way it's been implemented.

 

[Damn_Smooth]

and why do some fags spend their eves shitting on others dreams for a buzz wagwan wi that fanboys do my nut

Can anybody translate this for me?

 

[TheoneandonlyMrK]

I don't think tessellation is over rated, definitely not used in the right ways though.

IMO tessellation should be used when your right up close to something, mostly on things like bricks/details and things like that.

No need for it to smooth edges and stuff as that's not all that noticeable.

Also dynamic tessellation so you can smash a guys face in would be nice.

+1

id use it on terrain though also pantherx12
but not allways extreme tesselation ,it should be graded on the objects importance in the scene ie near stuff more detail far stuff a bit of tess not just random kerbs or 1 level of a game

essentially fanboys do my head in with their mindless waffle and debate damn smooth

 

[MxPhenom 216]

It is overrated, but on somethings i like what it does to the visuals.

 

[TheoneandonlyMrK]

i personally think its fine that a card vender will focus its pr on the good point of its card or to bend benchmark makers, their way to make their cards look good , i dont mind at all their tryin at least and something may yet come of it imho its just better to take all tech news with enthusiasm and a big pinch of salt im allways eager for every bit of tech,,,,, to be reviewed and properly then i can really get excited ,or not.

 

[SK-1]

Honestly mate, Unigine is a benchmark too. But you make a sweeping statement about not trusting any benchmarking utility out there, mate.

So what is it really mate? And besides mate, if you further check on Tech Report's revisions on their methodology, FPS isn't exactly the ultimate factor, mate.

Yup...you need some good old fashioned remedial sarcasm training...mate

 

[erocker]

Can anybody translate this for me?

+1. Wat?