HD6900 series to ship with 2GBs of Ram

[RussianSensation]

Another leaked slide:

Only 32 ROPs is a bit surprising though, but 2GBs GDDR5 would be nice to lead the industry. VLIW4 and 2x the geometry processing sounds juicy.

 

[Genx87]

I am assuming geometry processing means tesselation?
That wont close the gap with Nvidia. It also depends on the factor level. The 6800 series cards were upto 2x the tesselation performance at low factor loads compared to the 5000 series cards. Sounds like these might have the same anemic serial tesselator?

 

[SlowSpyder]

"VLIW4" seems more interesting... I guess that is what was mentioned earlier, that Cayman may be a different SP set up than Barts.

 

[RussianSensation]

Sounds like these might have the same anemic serial tesselator?

Not sure we can assume that just yet. The VLIW4 highlights that this isn't simply going to be a reduced Cypress chip like 6800 series was. Therefore, I don't think we can just assume that the tessellation engine will be the same.

2GBs of Ram and TDP < 300W is sending a message that this card is aiming to dominate 2560x1600 charts.

 

[Skurge]

I am assuming geometry processing means tesselation?
That wont close the gap with Nvidia. It also depends on the factor level. The 6800 series cards were upto 2x the tesselation performance at low factor loads compared to the 5000 series cards. Sounds like these might have the same anemic serial tesselator?

No, its not the same thing. its basically how many triangles in can process. Not sure if the measure per clock or at the cards given clock speed. Someone correct me If I'm wrong.

 

[Genx87]

Not sure we can assume that just yet. The VLIW4 highlights that this isnt' simply going to be a reduced Cypress chip like 6800 series was. Therefore, I don't think we can just assume that the tessellation engine will be the same.

2GBs of Ram and TDP < 300W is sending a message that this card is aiming to dominate 2560x1600 charts.

Well my assumption is based on the 6800 series also touting 2x the tesselation performance of the 5000 series. Which it does at sub 10 factors.

 

[Arkadrel]

"..snip..The kits come with a Cayman and some 6Ghz GDDR5, so the memory clocks will end up between 5.5-6GHz. ..snip.."

source: http://www.semiaccurate.com/2010/11/01/cayman-chips-turn-wild/

So: (I think this is how its calculated, please tell me If im wrong)

256bit bus width:
256x1.500 / 8 x 4 = 192 Gb/s memory bandwidth

384bit bus width:
384x1.500 / 8 x 4 = 288 Gb/s memory bandwidth (best case scenario with 6Ghz GDDR5)


For compairisons sake.. a Nvidia 480 has 177.4 Gb/s memory bandwidth (from looking at wiki).

the 288 Gb/s is the best case scenario (384bit bus width), if it is indeed true they will have GDDR5 at 6Ghz. They might be clocked lower than 6Ghz to save on power though... so it might be abit less.

 

[RussianSensation]

bit bus width:
384x1.500 / 8 x 4 = 288 Gb/s memory bandwidth

For compairisons sake.. a Nvidia 480 has 177.4 Gb/s memory bandwidth (from looking at wiki).

288 GB/s memory bandwidth would be a huge waste of $$$. ~ 200 GB/s would be more than enough. High end cards need more shaders and ROPs/TMUs before even thinking about 200+ GB/sec memory bandwidth.

 

[Skurge]

Not sure we can assume that just yet. The VLIW4 highlights that this isnt' simply going to be a reduced Cypress chip like 6800 series was. Therefore, I don't think we can just assume that the tessellation engine will be the same.

2GBs of Ram and TDP < 300W is sending a message that this card is aiming to dominate 2560x1600 charts.

The 32ROPs still troubles me.

 

[RussianSensation]

The 32ROPs still troubles me.

Agreed. Considering we got 5+ variations of HD6850/70 specs, you should still keep in mind that those specs are still just a rumour.

 

[Skurge]

"..snip..The kits come with a Cayman and some 6Ghz GDDR5, so the memory clocks will end up between 5.5-6GHz. ..snip.."

source: http://www.semiaccurate.com/2010/11/01/cayman-chips-turn-wild/

So: (I think this is how its calculated, please tell me If im wrong)

256bit bus width:
256x1.500 / 8 x 4 = 192 Gb/s memory bandwidth

384bit bus width:
384x1.500 / 8 x 4 = 288 Gb/s memory bandwidth (best case scenario with 6Ghz GDDR5)


For compairisons sake.. a Nvidia 480 has 177.4 Gb/s memory bandwidth (from looking at wiki).

the 288 Gb/s is the best case scenario (384bit bus width), if it is indeed true they will have GDDR5 at 6Ghz. They might be clocked lower than 6Ghz to save on power though... so it might be abit less.

wouldnt that mean the card would have something like 1.5 or 3GB of ram?

 

[Arkadrel]

@RussianSensation

Do you think its more likely they use a 256bit bus? +6Ghz chips to end up around 192 GB/s? and is that really enough?

wouldnt that mean the card would have something like 1.5 or 3GB of ram?

Thats true! if it was 384 it would have to be 1.536 GB of ram.

But the slide above states 2gb.
I guess that means if the slide is true, we ll see 256bit width on the new 6970's?

So fair assumption that it ll be 256bit and use 6Ghz GDDR5, and end up at 192 Gb/s bandwidth.

 

[nitromullet]

sounds pricey... We'll see soon enough how it performs.

 

[RussianSensation]

Do you think its more likely they use a 256bit bus? +6Ghz chips to end up around 192 GB/s? and is that really enough?

My 4890 had 128GB/sec bandwidth vs. 134GB/sec for the 470. The 470 is about 40-50&#37; faster on average. 4890 just proves that having too much bandwidth is wasteful (esp. when compared to HD5770). AMD would save a lot of $$$ by utilizing a 256-bit interface. Faster GDDR5 chips should guarantee more than sufficient memory bandwidth. They used the strategy of faster GDDR5 + 256-bit memory interface to compete against 320- & 384- bit chips from NV for several successful generations.

I don't see the point of increasing complexity of the PCB/GPU when their Cypress chips already worked very well with faster GDDR5 chips, unlike GF100 which struggles with faster GDDR5 (NV has an inferior memory controller at the moment - this is why they couldn't use faster GDDR5 and had to rely on a wider bus to compensate).

 

[bryanW1995]

yeah, amd can get the same bandwidth with 256 bit that nvidia gets with 384 due to their better memory controller.

Of course, after all of the bogus rumors about 68x0 I'm inclined to actually wait for launch and see real benches. I wonder if nvidia will send a gtx 480 lightning to anand for that review? ()

 

[ScorcherDarkly]

2GBs of Ram and TDP < 300W is sending a message that this card is aiming to dominate 2560x1600 charts.

I don't understand what TDP under 300W has to do with dominating anything. That fact by itself only means the card is power hungry. Correct me if I'm wrong, but 300W TDP is the limit for the PCI-e spec, so it has to be under 300. Are you trying to infer level of performance based off the power draw?

 

[Skurge]

I don't understand what TDP under 300W has to do with dominating anything. That fact by itself only means the card is power hungry. Correct me if I'm wrong, but 300W TDP is the limit for the PCI-e spec, so it has to be under 300. Are you trying to infer level of performance based off the power draw?

Yes, seeing as the last what? 3 Gens from AMD have used less than 200W for the highend single card. something over 225W is gonna have a nice increase in perf.

Does anyone have trouble believing that there is going to be a 6990 with 2 cayman chips?

 

[Ares1214]

32 ROP's is a little scary, unless they have changed things up on us to make "32" not as bad. But id guess that would be the significant bottleneck if it had one.

 

[dangerman1337]

I wonder what other stuff they've put into 69** that is seemingly making it considerably hotter and bigger than the 5870 without adding a 384 bit mem bus.

 

[RussianSensation]

Are you trying to infer level of performance based off the power draw?

AMD reports maximum TDP not average/typical TDP like NV does. So consider that HD6870 has a TDP of 151W, HD4890 a TDP of 190W, HD5870 has a TDP of 188W, and HD5970 a TDP of 300W. What do you think this tells you about an AMD card with a TDP between 225-300W? It's going to be very fast considering AMD delivers far more performance/watt than NV does.

For example, HD6870 with maximum TDP of 151W = GTX470 with TDP of 215W.

 

[busydude]

32 ROP's is a little scary, unless they have changed things up on us to make "32" not as bad. But id guess that would be the significant bottleneck if it had one.

I like to think AMD knows better than we do at the moment.

 

[Arkadrel]

What do you guys think about the 4D shader architecture (VLIW4). Will that just make the chip smaller in mm^2? or will that bring a little performance too (rumors say really good shader power)?


Rumors on [H]forum.
http://hardforum.com/showthread.php?t=1556495&page=3

they have:
~1920SP, 192GB / s, ~430 mm2,

 

[Ares1214]

What do you guys think about the 4D shader architecture (VLIW4). Will that just make the chip smaller in mm^2? or will that bring a little performance too (rumors say really good shader power)?


Rumors on [H]forum.
http://hardforum.com/showthread.php?t=1556495&page=3

they have:
~1920SP, 192GB / s, ~430 mm2,

430mm seems a bit big IMO. Id guess more along the lines of 380-400mm. And the VLIW4 doesnt necessarily "improve" performance, just makes things a lot more efficient so that they have the capacity to improve performance.

 

[Piano Man]

Arghh, I don't think there's anyway the 6950 is gonna be under 300 now

 

[Arkadrel]

32 ROP's is a little scary, unless they have changed things up on us to make "32" not as bad. But id guess that would be the significant bottleneck if it had one.

@[H]forum, a guy suggested that 5D shader architecture wasnt very effective... maybe 4D shaders means they need less ROPs?

Arghh, I don't think there's anyway the 6950 is gonna be under 300 now

I dont see a 6950 useing more power than a 480.
I see a 6970 useing more though (/_\) maybe its our turn to have nvidia guys pokeing at us for useing more power lmao. I just hope the 6970 is lightning fast, which I suspect it will be.

Also take it with a grain of salt... nvidia has odd ways of measureing its TDP (so it looks smaller than it is).

480 is listed as 250watts on Wiki, but people that measure it at the wall can get it up to 320watts.
When Nvidia list their TDP their not doing so under 100&#37; load showing the max value they draw.

ps. there are overclocked 480s that go over 400watts used. ^-^ As long as the 6970 isnt like that shouldnt be the end of the world.