AMD's REAL answer to GeForce GTX 480: AIB Custom 5970s

[Astennu]

I dont see this as a real answer because i still think a normal 5970 will be more then enough to beat the GTX480.

I think 5870 OC cards with 950 core clocks and 2GB mem can come quite close to the GTX480 performance.

But never the less its nice to see AMD will give us even more performance. I know i want to try one of those 5970 with 850 core and 2x2GB mem

That Arctic Accelero Xtreme looks good ! It was super on the HD4870 X2. I know i want one of those :P

[zalbard]

I dont see this as a real answer because i still think a normal 5970 will be more then enough to beat the GTX480.

It may be necessary if dual Fermi ever comes out, though.

[subaruwrc]

It may be necessary if dual Fermi ever comes out, though.

It won't because of the huge power consumption of the GF100 core.
If all the 512 CC are enabled, the TDP exceeds 300w, and we are talking about a single gpu...

[azza21]

The real answer with an unrealistic price. I will stick with my 4870X2 for now, unless i win the lottory.

[ReverendMaynard]

Limited production cards are far from "the answer". I guess I'm just getting bored with all of the hype surrounding both camps.

[SamHughe]

Limited production cards are far from "the answer". I guess I'm just getting bored with all of the hype surrounding both camps.

I'm gonna do something I haven't done for a loong time and write +1 for the post above.

[gOJDO]

It smells on MARS(read problems). How they gonna call it? SNICKERS or BOUNTY?

I call BS on the 28% better in 3DMV Extreme. I doubt it's going to be more than 15% better in average over the non-SNICKERS version.

[bill_d]

Limited production cards are far from "the answer". I guess I'm just getting bored with all of the hype surrounding both camps.

In difference to the Mars (Asus' dual GTX 285) the Ares won't be limited and is going to be produced and sold as long as there is a demand for the card

http://www.pcgameshardware.com/aid,7...nchmarks/News/

[Astennu]

It may be necessary if dual Fermi ever comes out, though.

In My Opinion thats a really big IF. With the currents rumors an pictures of the cooling, special cases (Elements V NV edition with special VGA airduct), Fan specs and the rumors about design + yield problems. I dont think we will see a dual fermi soon.

And if they it might even be slower then 2x GTX470 in CF. Because they need to disable more SP's and lower the clocks to not exceed the 400 watt.

I guess they will have to wait for: 1: To fix 40nm and improve the design 2: 28nm.

It smells on MARS(read problems). How they gonna call it? SNICKERS or BOUNTY?

I call BS on the 28% better in 3DMV Extreme. I doubt it's going to be more than 15% better in average over the non-SNICKERS version.

Why ?

The Clockspeeds are 20% higher. And memory is 100% bigger. You dont know what the impact of the memory will be.

Maybe its the same as from the Radeon 8500 64MB > Radeon 8500 128MB i saw huge boosts in performance in all 3Dmark 2001 SE tests.

And a more recent example HD4870 512 vs 1024. Average 15% performance boost. Starting at 1280x1024 with a 10% boost.

So it might be possible depending on the resolution and the arcitecture of the RV870. Maybe it can utilize more memory capacity better.

So i think its possible to see 28% performance increase in benshmarks and games.

[gOJDO]

Why?


The Clockspeeds are 20% higher.

And memory is 100% bigger. You dont know what the impact of the memory will be.

Maybe its the same as from the Radeon 8500 64MB > Radeon 8500 128MB i saw huge boosts in performance in all 3Dmark 2001 SE tests.

And a more recent example HD4870 512 vs 1024. Average 15% performance boost. Starting at 1280x1024 with a 10% boost.

So it might be possible depending on the resolution and the arcitecture of the RV870. Maybe it can utilize more memory capacity better.

So i think its possible to see 28% performance increase in benshmarks and games.

1) The clocks are not 20% higher. PCI-e, which is the interface connecting the two GPUs and the NB is clocked same. The core clock is clocked 17% higher.

2) You can count on one hand the games that in a very specific case(at certain scene when playing with everthing @max: resoultion, settings, AA and AF) can benefit from frame buffer larger than 1GB. There is no game yet that can utilze 2GB frame buffer at everything maxed out. I bet that in 95% of the games there would be no difference in perofrmance between the 4GB and the 2GB version of the card @same clocks.

3) Never compare single GPU scalling with multi GPU scalling. Just check the difference in performance between Asus MARS and regular GTX 295. MARS has more theoretical computing power compared to the regular GTX 295, but in reallity the difference was marginal.

4) I owned a systems with a 4870X2 on a Q9450@3.9GHz, and 2x4890, a GTX295, a 5870 and 2x5870 on an i7-920 @4.2GHz~4.4GHz. I did a lot of benchmarks and came to the conclusion that a dual-GPU performance never scales linear with frequencies, although the CPU bottleneck was basically avoided.

So, that 28% gain is a cherry picked result form a lot of benchmarks. And if that is the best, I guess the average will be at half of that at best.

[Chickenfeed]

I apologize if this has already been mentioned. There was some interesting info at the end of the article which seems to support the title of this thread....

Saphire's custom 4GB overclocked 5970...here.
Same clocks as the Asus but only 2 8pins. The board itself is smaller however (reference size)meaning it won't have potential clearence issues like the Asus ( which is a taller than reference pcb )

Their comment at the end is very interesting if true ( not to mention somewhat sad )

Officially, this product will be named just as every other overclocked Radeon HD 5970 but in reality, we're talking about "HD5990", AMD's double whammy to respond to nVidia's GeForce GTX 480. Given that we managed to learn partner allocations for GTX 470 and GTX 480, it is not surprising to see AMD lifting the lid on the overclocked parts. According to our information, partners plan to compete against nVidia on 1:1 ratio between overclocked 5970 boards and GTX 480, which is a pretty interesting plan. Gotta love the competition, right?

This might just be a premature comment on their part however if it is greeted with legitimacy... custom 5970s 1:1 with 480s? I take away two things from this, the first being less likely. 1) The 480 may compete more closely ( in price at least ; ideally performance as well ) to the 5970. 2) Availability will be as bad as rumored. Given the fact that the 5970 is already not an overly available part... custom overclocked premium models will be ever more so. That said though, I don't see AMD sitting ideally content with their partners designing higher performance solutions on their lonesome. We should be seeing a higher performance reference part sooner or later... ( Would love to see a 2GB 5890 at 1000/1300... )

[Astennu]

1) The clocks are not 20% higher. PCI-e, which is the interface connecting the two GPUs and the NB is clocked same. The core clock is clocked 17% higher.

2) You can count on one hand the games that in a very specific case(at certain scene when playing with everthing @max: resoultion, settings, AA and AF) can benefit from frame buffer larger than 1GB. There is no game yet that can utilze 2GB frame buffer at everything maxed out. I bet that in 95% of the games there would be no difference in perofrmance between the 4GB and the 2GB version of the card @same clocks.

3) Never compare single GPU scalling with multi GPU scalling. Just check the difference in performance between Asus MARS and regular GTX 295. MARS has more theoretical computing power compared to the regular GTX 295, but in reallity the difference was marginal.

4) I owned a systems with a 4870X2 on a Q9450@3.9GHz, and 2x4890, a GTX295, a 5870 and 2x5870 on an i7-920 @4.2GHz~4.4GHz. I did a lot of benchmarks and came to the conclusion that a dual-GPU performance never scales linear with frequencies, although the CPU bottleneck was basically avoided.

So, that 28% gain is a cherry picked result form a lot of benchmarks. And if that is the best, I guess the average will be at half of that at best.

1: Who says the card is PCI-E limited ? And the Memory clock is 20% higher. Core clock is 17%.

2: I dont think 3dmark vantage scores are very specific. Normally 3Dmark does not scale well with higher clocks. I know normal scaling is not 100% with clocks 12% higher clocks mostly result in 8-9% higher fps. And i know there are no games that will use 2GB. Keep in mind the 5970 classic is not really 1GB !!. Because its a dual GPU card. CF and SLI in there current form dont have a shared memory pool. Every GPU has one GB. But because they both need to render the same scenes (only different frames) they both need the same information in there memory buffer. And because of that CF and SLI setups of 2gb effectively have 1GB.

So the 4GB effectively is 2GB. And there are no games that will use all of that at the moment but i think there will be games that go beyond 1GB. For one GTA4 needs more then 1GB to run maxed out. And i bet there are modern games that will show nice boosts when going from 1 to 2GB. But just like the transition from 512 to 1024 there will also be games that show almost no performance gain.

And i think you will see some difference between the 5970 2gb and the 4gb. I guess you could see 5-10% depending on the resolution.

and also do not forget eyefinity. Its possible to get resolutions up to 24 mega pixel. 6x more then a single 30" display. And thats the most important reason for the bigger frame buffer. I think vs dual screen there are quite a lot of people using 3 screen eyefinity with 3x22" you already have 5.7 mpixel vs 4.09 of a single 30" display. And yes i know people who use it do play games.

3: I'm not comparing single to multi gpu scaling i'm comparing a dual 5970 vs another dual 5970. Both multi GPU. I know there is a potential bigger lower gain when overclocking. But when you have GPU limited situation SLI and CF scaling are quite good (in modern games). I dont know about the mars but i know CF OC did help a lot for the performance when you where GPU limited. But there are not a lot of games that will max out a 5870 CF setup. Only really new games like Dirt II on DX11 mode can. Or when running high resolutions 2560x1600+.

Your not buying cards like these to play on 1280x1024. A single 5850 is more then enough for 1920x120 (2.4 Megapixel) but a 5870 is still a bit to slow for 2560x1600 in some more demanding games (24 Fps when running STO)

4: i agree with that. It also depends on the game and if it has good CF/SLI support. If it has good support you can see the scaling up to 95%.

And your GPU OC will scale just as well.

But if you meant with your reaction that games with bad CF/SLI wont scale well with the higher clocks you are correct. But when you compare it with a normal 5970 i think that i overall a 20%+ gain is possible. Because you compare dual vs dual and if CF is not working you will still have the clockspeed and memory buffer advantage. (deepening on the resolution because you need the gain of the 2gb per gpu buffer) on lower resolutions i guess it will be more like 15% gain.

But we will just have to wait and see how it will perform in the real world. There are already 3 vendors that will bring a 850-1200 4GB 5970 to the market. So i think we will some nice reviews when that time comes.

[Aerwidh]

It smells on MARS(read problems). How they gonna call it? SNICKERS or BOUNTY?

I call BS on the 28% better in 3DMV Extreme. I doubt it's going to be more than 15% better in average over the non-SNICKERS version.

That gave me a good laugh. Thank you

[zalbard]

It won't because of the huge power consumption of the GF100 core.
If all the 512 CC are enabled, the TDP exceeds 300w, and we are talking about a single gpu...

Have you not seen the card from the OP? Who cares about 300W in such extreme cases?

[gOJDO]

1: Who says the card is PCI-E limited ?

I do. Grab an IDE disk(and disconect every SATA drive you have), push the PCI-e clock to the max and do some benches. Then bring the PCI-e clock to stock and do the same beches again. Let us know the difference.

2: I dont think 3dmark vantage scores are very specific.

3D Mark is almost useless benchmark for measuring real-world gaming performance. It can be used only as a meter when you compare two cards with same architecture, but with different clocks on the same CPU. That's why I call BS on the 28% better part and why I doubt that the tweakedd 5970 will be more than 15% faster than the stock in average.

Normally 3Dmark does not scale well with higher clocks. I know normal scaling is not 100% with clocks 12% higher clocks mostly result in 8-9% higher fps.

Different types of graphics cards act differently. If the VGA has plenty of VRAM bandwidth then upping the VRAM clocks will add no points to the result, but if the VGA is bandwidth hungry you will se almost linear scalling in performance with only increasing the VRAM clock. The best example for that is the 8800GT. It's bandwidth starved.

Keep in mind the 5970 classic is not really 1GB !!.

Theoretically it has 2GB on board, but in reallity it is exactly a 1GB card.

Because its a dual GPU card. CF and SLI in there current form dont have a shared memory pool. Every GPU has one GB. But because they both need to render the same scenes (only different frames) they both need the same information in there memory buffer. And because of that CF and SLI setups of 2gb effectively have 1GB.

I perfectly understand that.

So the 4GB effectively is 2GB.

That's why I was talking about 1GB vs 2GB frame buffer(or reference 5970 vs tweaked 5970).

And i think you will see some difference between the 5970 2gb and the 4gb. I guess you could see 5-10% depending on the resolution.

I don't think so. IMO there would be no gain at all in 99% of the cases.

and also do not forget eyefinity. Its possible to get resolutions up to 24 mega pixel. 6x more then a single 30" display. And thats the most important reason for the bigger frame buffer. I think vs dual screen there are quite a lot of people using 3 screen eyefinity with 3x22" you already have 5.7 mpixel vs 4.09 of a single 30" display. And yes i know people who use it do play games.

BINGO! I forgot eyefinity There these tweaked 5970's are going to shine! So, it will be the card for those who will buy a pair and at least 3 30" monitors.

3: I'm not comparing single to multi gpu scaling i'm comparing a dual 5970 vs another dual 5970.

OK, because you menitoned single Radeon 8500 64MB and single Radeon 4850 512MB. These cards have nothing common with a card like the 5970(the stock and the tweaked).

But we will just have to wait and see how it will perform in the real world. There are already 3 vendors that will bring a 850-1200 4GB 5970 to the market. So i think we will some nice reviews when that time comes.

Yeap. We'll wait and see.

[ReverendMaynard]

http://www.pcgameshardware.com/aid,7...nchmarks/News/

They are not going to be running huge production on that card, there isn't and never will be a huge market for upper echelon cards like that.

[Astennu]

I do. Grab an IDE disk(and disconect every SATA drive you have), push the PCI-e clock to the max and do some benches. Then bring the PCI-e clock to stock and do the same beches again. Let us know the difference.

Ok there will be a difference but i think it wont be so big. Also see this test: http://www.techpowerup.com/reviews/A...press_Scaling/

Its a Single 5870. Do do see some difference but its not major when moving from 8x to 16x. So 8x per GPU should be enough.

3D Mark is almost useless benchmark for measuring real-world gaming performance. It can be used only as a meter when you compare two cards with same architecture, but with different clocks on the same CPU. That's why I call BS on the 28% better part and why I doubt that the tweakedd 5970 will be more than 15% faster than the stock in average.

ok i totally agree there 3dmark is crap and is nothing close to the real world

Different types of graphics cards act differently. If the VGA has plenty of VRAM bandwidth then upping the VRAM clocks will add no points to the result, but if the VGA is bandwidth hungry you will se almost linear scalling in performance with only increasing the VRAM clock. The best example for that is the 8800GT. It's bandwidth starved.

That is true. But the case of the 5870 and 5970 its quite balanced. I did some tests with 850 MHz core and only 1000 MHz mem. Then you see a drop. But when you have 750 Core and you upp the memory from 1000 to 1200 the boost is really small. And with the 5970 OC they increase both almost 20%.

BTW: the 4850 is also bandwidth starved

That's why I was talking about 1GB vs 2GB frame buffer(or reference 5970 vs tweaked 5970).

I don't think so. IMO there would be no gain at all in 99% of the cases.

OK, because you menitoned single Radeon 8500 64MB and single Radeon 4850 512MB. These cards have nothing common with a card like the 5970(the stock and the tweaked).

I know the architectures are different and there i a chance you will not notice anything. The HD4870 1024 was 15% faster then the 512mb. But the 4850 on the other hand was not faster at all. It did had halve the bandwidth of the 4870 that might be the cause. But as far as we know the 5870 had enough bandwidth. Core speed boosts do a bit more then mem speed boosts. So far i have not found any card in the HD57 tm 59 series that is memory bandwidth starved.

And i dont say it will be 25% faster but i wont be surprised if i is looking at the past there are multiple cases where a bigger framebuffer did give a nice overall boost. I for one never expected that the 4870 1024 would be 10% faster on 1280x1024. I thought you would only notice it at 1920+. But this was not the case.

[Mungri]

I suggest a new plan for ATI. Dont release any new 5000 series cards, but focus on pushing out their next gen DX11 cards A.S.A.P.

[gOJDO]

I suggest a new plan for ATI. Dont release any new 5000 series cards, but focus on pushing out their next gen DX11 cards A.S.A.P.

With no competition and still not paid off(I'm talking about the money invested for research, development and marketing) Radeon 5000 series, that would be the smartest plan if their goal is to bankrupt.

[zalbard]

I suggest a new plan for ATI. Dont release any new 5000 series cards, but focus on pushing out their next gen DX11 cards A.S.A.P.

I think they are already working on it...

With no competition and still not paid off(I'm talking about the money invested for research, development and marketing) Radeon 5000 series, that would be the smartest plan if their goal is to bankrupt.

I doubt making 5xxx took really so much resources, they used nearly the same arch as before, plus tessellation is nothing new for ATI.
They have enough 5xxx chips out already (if not too many). Time for something new.

[Astennu]

I suggest a new plan for ATI. Dont release any new 5000 series cards, but focus on pushing out their next gen DX11 cards A.S.A.P.

I seems they are. so far there is no indication of a RV890. We know they will allow HD5870 OC moddels with 900-950 clocks and also 5970 with 850 clocks.

Depending on the performance of the GTX480. That boost could be enough to stay competitive.

I think (if there are no unexpected problems) AMD Will be able to release the HD6xxx series this year.