Radeon HD 7970 Raw Specs Leaked

[TheMailMan78]

If slide leaked on Orb-Hardware is to be believed we GPU consumers are in for a pretty big treat in the next month or so. The slide shown below states that the AMD 7970 will have a default core clock speed of 925MHz and a whopping 3 GB of GDDR5 memory.

It also sports a 3.5 TFLOPs precision floating point. Which would put it well beyond the NVIDIA flagship single GPU solution. The slide states the a ROP count of 32, against an earlier speculated count of 48. This could be because AMD may have delinked ROP clusters from memory bus. The cooler itself is under the trademark AMD black shroud so there is no way to see if it uses the rumored "vapor chamber" as seen in after market solutions.

Looking past all the beastly prowess of this slide one cannot help but think about power draw. The "leaked" slide states the 7970 will have a peak power draw of 300 W and an idle draw of 3 W. We will have to wait for review to see if any of these amazing stats are true.



View at TechPowerUp Main Site

 

[ViperXTR]

32 ROPs on a 384bit 3GB GDDR5 ram '__'

 

[Maban]

Donanimhaber has a few more slides.

I'm a little disappointed with only 32 ROPS but I'm quite interested in GCN.

 

[Lionheart]

32ROPS wtf, should still be a beast of a card

 

[WarraWarra]

Idle 3w sure it is not 30w surely this is not a mobile gpu as I suggested they do and if it is to achieve the 3w minm then 300w is way over the top for the top end.

ATTENTION: It seems this slide is legit but little bit old, GPU clocks and Load TDP are on real cards lower! GPU clocks are probably under 900 MHz and Load TDP +/- 200W.

 

[Volkszorn88]

AMD, hurry up and rape my wallet right f'ing now!

 

[claylomax]

Donanimhaber has a few more slides.

I'm a little disappointed with only 32 ROPS but I'm quite interested in GCN.

Thank you very much for the link, more slides indeed.

 

[63jax]

...so no Rambus Ram? it would've been nice.

 

[Delta6326]

If 3W is possible it sounds like they have found out how to do more than lower clocks and Volts.

Over 3TFLOPs! I really need to upgrade my 2x 4850 now that I see these specs!

 

[Zubasa]

32ROPs = Kepler here I come.

 

[Sihastru]

Remember how Cayman was a "Dual Graphics Engine" GPU? It was essentially two Barts GPUs glued together, each with dual 64-bit dual channel memory controllers, where the 2 GB of RAM came from.

This new GPU looks more like a "Tri Graphics Engine" GPU (they gave it a fancy marketing name, GCN = "Graphics Core Next"), so essentially three Barts GPUs glued together (albeit at 28nm), each with dual 64-bit dual channel memory controllers, and this is why it has exactly 3GB of RAM.

Conclusion?

Anyone with a tri-fire HD6870 setup can tell us exactly how this HD7970 card will perform.

"Three HD6870s on a stick" (you can quote me on that)

 

[gorg_graggel]

just because the quantity of ROP is the same as the previous generation, doesn't mean their quality will be the same...
for all we know, they could have beefed them up to be double as fast...

or maybe it just didn't improve performance enough to justify the extra die space...
after all they have been pretty beefy to start with...

 

[Zubasa]

just because the quantity of ROP is the same as the previous generation

Make that the generation before that previous generation as well.

 

[repman244]

Make that the generation before that previous generation as well.

It still doesn't mean that they are the same, it don't think you can just compare the numbers with previous generation.

 

[gorg_graggel]

It still doesn't mean that they are the same, it don't think you can just compare the numbers with previous generation.

^^this

 

[Zubasa]

It still doesn't mean that they are the same, it don't think you can just compare the numbers with previous generation.

^^this

I never said it is the same, for all we know there is a company that can make things worst than their previous generation.
Even if they some how manage to to make the ROPs twice as powerful, top-end Kepler has exactly 64ROPs and that is counting on nVidia some how doesn't know how to make their GPUs better.

 

[repman244]

I never said it is the same, for all we know there is a company that can make things worst than their previous generation.
Even if they some how manage to to make the ROPs twice as powerful, top-end Kepler has exactly 64ROPs and that is counting on nVidia some how doesn't know how to make their GPUs better.

That is true

But we still don't have anything about Kepler (at least not as much we know about HD7xxx) so we can only guess how Kepler will perform(and we could also say the same for HD7xxx).
Guessing the performance based on paper numbers can backfire (a.k.a. Bulldozer )

 

[nt300]

AMD Radeon HD 7970 Launch Pulled Forward to 22nd December, Rumour Claims

Tell me what is the benefit for increasing the number of (ROPs) Render Output Units
Or is it (ROPs) Raster Operations Or what ever let call them ROPs

6 more days till we see the monsterous HD 7970.

AMD Radeon HD 7970 Launch Pulled Forward to 22nd December, Rumour Claims

http://www.itproportal.com/2011/12/16/amd-radeon-hd-7970-launch-pulled-forward-22nd-december-rumour-claims/

 

[gorg_graggel]

ROP = Raster Operation Processor

it's the last step in a graphics pipeline, where all data is put together to be copied into the framebuffer, which then gets displayed on the monitor...anti aliasing and depth information is also processed there... (overly simplified)


so increasing the number of those units would only benefit, when 1) the units have not been improved over the last generation and to get more performance you need to have more ROPs, or 2) they have not been fast enough already...

 

[TheMailMan78]

Many 1337 thanks to DanishDevil on the scoop.

 

[phanbuey]

Performance is not going to be HUGE since its a totally new approach, which usually means no 2x performance jump (again totally speculation take with a lethal dose of salt.). Some games I am sure the 6970 will be close to the speed, and in others it will get blown away.

Also, dont forget about AMD's infamously fast driver support. They probably only optimized 3Dmark and Unigine for now, and Crysis 2 and other games are still rendering with the "VGA adapter driver" from windows 95.

 

[TheMailMan78]

Performance is not going to be HUGE since its a totally new approach, which usually means no 2x performance jump. Some games I am sure the 6970 will be close to the speed, and in others it will get blown away.

Also, dont forget about AMD's infamously fast driver support. They probably only optimized 3Dmark and Unigine for now, and Crysis 2 and other games are still rendering with the "VGA adapter driver" from windows 95.

The last time they changed architecture it was from the 48xx series to the 58xx series. If that wasnt a huge jump in performance I dunno what is. The 69xx series is just a refined 58xx series so don't base your opinions off of that.

 

[phanbuey]

The last time they changed architecture it was from the 48xx series to the 58xx series. If that wasnt a huge jump in performance I dunno what is. The 69xx series is just a refined 58xx series.

I agree that was huge - but that wasnt a new Architecture IMO- it was the same VLIW concept tweaked. and in these cases performance increase is massive and proportional.

4870 had 800 stream procs... 5770 also had 800 tweaked with DX11 instead of DX10.1, and performance was similar.
5870 had 1600 stream procs - was double the speed of the 4870 give or take.

2900 - 3800 -4800 - 5800 - 6900 are all general variants of the same concept, 7xxx series seems a bit more radical, uses a totally new approach - closer to the 1950XTX to 2900XT shift (again, take with a lethal dose of salt purely my 2c)

 

[FierceRed]

Games this GPU needs to be benchmarked against at 1920x1200 & 2560x1600 before people can justifiably start to whine about the 32 ROPs:

• Batman: Arkham City under DX11
• Battlefield 3 on Ultra Preset, no custom bullshit
• Metro: 2033 because its the new Crysis
• Metro: Last Light upon release
• Skyrim because people will froth if it isn't benchmarked... and then they'll froth anyway.
• Witcher 2 with and without Ubersampling

And by the way, unless all the major sites have had 7970s since the beginning of December, I doubt we're going to see "reviews" including benchmarks next Thursday. Which is fine because I wouldn't want to see a 7970 benchmarked on pre-12.1 drivers anyway.

Wouldn't mind seeing these benchmarks on 120Hz monitors as well, now that I think about it. This vSync idiocy needs to end.

 

[TheMailMan78]

I agree that was huge - but that wasnt a new Architecture IMO- it was the same VLIW concept tweaked. and in these cases performance increase is massive and proportional.

4870 had 800 stream procs... 5770 also had 800 tweaked with DX11 instead of DX10.1, and performance was similar.
5870 had 1600 stream procs - was double the speed of the 4870 give or take.

2900 - 3800 -4800 - 5800 - 6900 are all general variants of the same concept, 7xxx series seems a bit more radical, uses a totally new approach - closer to the 1950XTX to 2900XT shift (again, take with a lethal dose of salt purely my 2c)

The 7970 is rumored to have 128 texture units, and 2,048 stream processors.