Let the 6970 Rumors Begin

Status
Not open for further replies.
Pretty interesting. I'll see what Nvidia comes up with and what kind of audio spec they go with. The fact that I'm now using my video card as an audio source now weighs heavily on my decision, too.
 
so no 28nm for ATI or Nvidia until late 2011?...I would love to upgrade my 5870 but not for a 20% improvement...the Tesselation improvement is purely for the Heaven benchmark as no games are in the pipeline that need all that Tesselation power (only one using it currently is Metro 2033 I believe)

hope 28nm will be available sooner
 
If that slide is correct, it looks pretty weak. That is the same ROP/Z-Stencil as the 6870, likely even still a 256 bit bus.
 
Snowdog said:
If that slide is correct, it looks pretty weak. That is the same ROP/Z-Stencil as the 6870, likely even still a 256 bit bus.
Click to expand...

Doesn't matter if its being used efficiently. VLIW4 means new shader architecture, and look at the size of that chip :eek:
 
ZerazaX said:
Doesn't matter if its being used efficiently. VLIW4 means new shader architecture, and look at the size of that chip :eek:
Click to expand...

Die size cant be seen in that pic. Thats just the pin out grid. Die could still be smaller, but I agree with your other ppoint about Vliw4. I have very high expectations on Vliw4.
 
Snowdog said:
If that slide is correct, it looks pretty weak. That is the same ROP/Z-Stencil as the 6870, likely even still a 256 bit bus.
Click to expand...

Eh? most people looking at this leak think this is gonna be beastly. Chip is huge. I know we can't really see the exact size but it coincides with rumors we know its not the same size as 5870. 2gb is a plus and new architecture as well

edit: rumored to have 1920 SP if this is true bye bye 'Fermi done right' will be eating dust.
 
Last edited:
Vexerz said:
Eh? most people looking at this leak think this is gonna be beastly. Chip is huge. I know we can't really see the exact size but it coincides with rumors we know its not the same size as 5870. 2gb is a plus and new architecture as well

edit: rumored to have 1920 SP if this is true bye bye 'Fermi done right' will be eating dust.
Click to expand...

The Chip isn't shown, you can't tell the chips size from a schematic showing an outline of the package.

I am judging by the slide that shows ROPs/Stencil counts have not incresed at all. Right there that will bottleneck in many situations.

Next up with the ROP count and 2GB memory, I am betting that this still has 256bit memory bus, yet another big bottleneck.

A nice shader boost, but with no more ROP/memory width, there will be too much bottlenecks to bother with a monster shader count. I suspect it will be 1680 SPs or less.

If that slide is correct,this is going to be a fair bit weaker than people are expecting.
 
Snowdog said:
The Chip isn't shown, you can't tell the chips size from a schematic showing an outline of the package.

I am judging by the slide that shows ROPs/Stencil counts have not incresed at all. Right there that will bottleneck in many situations.

Next up with the ROP count and 2GB memory, I am betting that this still has 256bit memory bus, yet another big bottleneck.

A nice shader boost, but with no more ROP/memory width, there will be too much bottlenecks to bother with a monster shader count. I suspect it will be 1680 SPs or less.

If that slide is correct,this is going to be a fair bit weaker than people are expecting.
Click to expand...
Let me ask how do you factor in VLIW4 into this?
 
Vexerz said:
Let me ask how do you factor in VLIW4 into this?
Click to expand...

I don't.

I expect the card will have excellent shader power, more than ample for it's bottlenecks from the ROPs/Memory bus.

But I expect the 580 (AKA the fixed 548) with its 384 bit bus will be faster in most situations.
 
well, seeing as the 57xx series was not bottlenecked by 128bit memory bus, and 58xx series as well as 68xx series dont seem to be coming too close to that bottleneck- how do you know it will hold back the 69xx? ATI/AMD has been praised for their memory controllers. You are also not taking into account ho fast that GDDR5 will be- They could be putting super fast GDDR5 chips on there and getting plenty of bandwidth not to bottleneck. You are just hoping that this is the case. I imagine that if those were big bottlenecks for them, they would have changed them. Lets wait and see :)
 
Robin B said:
Let the leaks flow ...



Uploaded with ImageShack.us



Uploaded with ImageShack.us

http://www.chiphell.com/thread-133368-1-1.html
Click to expand...


From the OP in chiphell:
PHK: informant sent me a NV sent from the Email, the letter referred to knock down the Cayman is not a problem 580. But the letter did not say the Cayman in the end is Pro or XT, but my guess is XT. The letter also nominal GTX 580 faster than 480 in the game 20 to 30%. Finally NV said Cayman's performance should not reach level of GTX580
Click to expand...

If true...
- GTX 580 is 20 to 30% faster than GTX 480, but in what game though?
- Cayman (XT or Pro?) can't reach the level of performance of a GTX 580

Other interesting quotes from that thread: (Rough Google translations.)

From 903347749
... I am optimistic about this generation of the 6900 series and 6800 series...
I saw the listing of AMD's performance per square millimeter, regardless of the large size, the cost is essentially on NV... what NV should do is improve the performance per square millimeter or be one day defeated by AMD.
Click to expand...

From shiori_2003
NV's stock looks "empty"
Huang points out the wind had to divert attention
Click to expand...

From nakedsoul
I do hope the GTX 580 is 20-30% faster than the 480 while full load power is down by 30%...

Otherwise that's wishful thinking
Click to expand...

From allensakura
It seems really 4D a, 1920SP / 4 = 480US is Bart 1120 / 5 = 224US more than doubled

However, with Bart as ROP or 32, even if the MC faster bandwidth of up to 192GB / s or a bad way, the core frequency to the force that must be
Core size should be slightly more than R600, the level reached 430 mm2
This specification can not be better than 5970 is more strange

In accordance with the formula for calculating power, the TDP in 6870 as a standard, the same voltage and frequency settings, power consumption can be drawn about 6,970 for the 254W
Click to expand...

This was interesting, he believes it'll be twice the cost of Barts (6800-series) if going by the SP count. It roughly translates to: "Even if the bandwidth is up to 192 GB/s or lower, the core frequency to force that must be that the core size must slightly larger than the R600, which is 430 mm2. This specification can't be better than the 5970 which is more strange."

He also thinks the power consumption for the 6970 might be 254W if using the TDP from the 6870 as a standard to calculate from. This guy even goes on to say that if the GTX 580 is 384-bit, the memory bandwidth should be a tie.

It's even funny that Charlie from SA is mentioned... lol

Wow, he's even known in China.
 
octoberasian said:
Interesting, second site to mention 255W. The poster from Chiphell even suggested something close-- 254W-- for the the card. And, if the performance is true in Unigine, that's pretty impressive numbers at that setting.
Click to expand...

Well, honestly, this card doesn't have to beat the 580. All it has to do is be cheaper and only marginally slower. We all know that the 580 won't (likely) retail for anything less than $500. The original 5870 retailed for around $400.
 
I don't understand the memory bus argument. I don't believe the 5870 was bandwidth limited, and a simple clock speed increase over the 5800 series will generate more bandwidth. Most 5870s can run up to 1300MHz GDDR5 when overclocked, generating 5.2GHz of performance, while NVIDIA's memory controller typically has not been up to snuff to run speeds that high. That would be my guess as to why they are using a 384-bit architecture. It's more cost effective to stay with 256-bit in terms of board costs.
 
The slide says 6970 and 6950. With low ROP count this might be a 6950 stencil we are looking at.

And Snowdog not factoring the new architecture is huge and how it works with the rest is just jumping the gun to drawn any good speculation.
Personally I'm guessing 25%-30% > 480 performance and double tessellation over the 5870.


(btw why aren't the img tags working in this forum)
 
On a side-note, I was planning on skipping this generation as I have 2x watercooled 5870s, but if the performance increase in Eyefinity is noticeable then I would consider an upgrade. The 2GB of memory standard is a big plus if the shader performance is there.
 
Vexerz said:
The slide says 6970 and 6950. With low ROP count this might be a 6950 stencil we are looking at.



(btw why aren't the img tags working in this forum)
Click to expand...
The slide lists separate power consumption figures so I would assume they would have listed the # of ROPs separately if that was the case.
 
Rizen said:
The slide lists separate power consumption figures so I would assume they would have listed the # of ROPs separately if that was the case.
Click to expand...

That's Max board power that AMD likes to use but look at the title it lists both 6970/6950. They just didn't list anything else for both except MBP
 
Rizen said:
I don't understand the memory bus argument. I don't believe the 5870 was bandwidth limited, and a simple clock speed increase over the 5800 series will generate more bandwidth. Most 5870s can run up to 1300MHz GDDR5 when overclocked, generating 5.2GHz of performance, while NVIDIA's memory controller typically has not been up to snuff to run speeds that high. That would be my guess as to why they are using a 384-bit architecture. It's more cost effective to stay with 256-bit in terms of board costs.
Click to expand...

I was expecting increases across the board; more SPs, ROPs, Memory Bus. If they are only increasing SPs, I believe that will limit performance gains in some situaltions. Time will tell.

I was initially thinking it would > NV 580, now that seems doubtful.

But I think AMD will keep bang/buck.
 
82221693.jpg


94189504.jpg


http://www.chiphell.com/thread-133368-1-1.html
 
Snowdog said:
I am judging by the slide that shows ROPs/Stencil counts have not incresed at all. Right there that will bottleneck in many situations.
Click to expand...

A drop in ROPs didn't hurt the 6870/6850 that much so I don't see how you can make that assumption.

Next up with the ROP count and 2GB memory, I am betting that this still has 256bit memory bus, yet another big bottleneck.
Click to expand...
The 5870 is only 10% slower than the GTX480 with a 256bit bus...Besides, you have no idea what the speed of the memory they're going to use will be, don't make assumptions without facts. This is a fact, AMD has always been ahead of Nvidia in using newer and faster types of memory.

A nice shader boost, but with no more ROP/memory width, there will be too much bottlenecks to bother with a monster shader count. I suspect it will be 1680 SPs or less.
Click to expand...
More baseless assumptions. You have zero details about the underlying vliw4 architecture, SPU utilization improvements, you don't know if caches have been widened. Again, don't make assumptions without facts.

If that slide is correct,this is going to be a fair bit weaker than people are expecting.
Click to expand...
If that slide is correct, it tell us not a whole lot about the new chips.
 
Last edited:
Seems like both NV & AMD's new cards are seeming less stellar just before release. Until we know more as Lorien states then we wont know how it translates to real world gaming. This isn't a new architecture after all. A refresh should be a performance improvement over the last and BOTH appear to be on the right track.

Maybe not earth shattering but both will probably be around +30% over last gen with optimized drivers.
 
Lord_Exodia said:
Maybe not earth shattering but both will probably be around +30% over last gen with optimized drivers.
Click to expand...

I agree, but that means the order doesn't change, NV keeps most powerful GPU.


Lorien said:
A drop in ROPs didn't hurt the 6870 that much so I don't see how you can make that assumption.
Click to expand...

The drop in ROPs was partially offset by an increase in clock speed and it still lowered performance vs the 5870, so it easy to see how I made that assumption.

Everything is speculation right now. If that slide is reality it shows a lot of units holding steady and that will definitely put a cap on performance in some situations.

I am not saying performance won't improve, just that it won't improve as much as it would with 50% more ROPS, and 320 bit bus.
 
why would it need more ROPs =)? ROPS are not as important as they used to be, the amount of pixel shaders is not longer 1:1 with ROP.
 
Digital Viper-X- said:
why would it need more ROPs =)? ROPS are not as important as they used to be, the amount of pixel shaders is not longer 1:1 with ROP.
Click to expand...

I guess he's thinking that by increasing the Shader count and not increasing the ROPs you will bottleneck gains in some instances.

Correct me if I'm wrong Snowdog but are you thinking that the extra Shaders will be a waste at times due to a lack of ROPs, and only 256 bit memory interface, and at times it will impact the performance?

It sounds a little logical but as Lorien said we dont know what speed gddr AMD is using and there could be a massive memory bandwidth boost which could open up the bottleneck (if any) a little more, theoretically of course
 
Snowdog said:
I am not saying performance won't improve, just that it won't improve as much as it would with 50% more ROPS, and 320 bit bus.
Click to expand...
That's the thing, you don't need a brute force architecture approach to increase performance significantly. Focusing on efficiency can bring about the same increases if done right. We know far far less about Cayman than we knew about Barts a month before launch. But it sure is fun to speculate lol.

I'm also thinking the relatively low ROP count is due to 40nm. Remember this chip design was supposed to be implemented on the 32nm node and is now being backported to 40nm, so some sacrifices had to be made. Imagine this architecture in 28nm, transistor density would skyrocket.
 
Last edited:
Ok fun time speculation:
VLIW5 (Very Long Instruction Word) from existing 5 series seen here to the 4 which is what the AMD slide is showing.

16SP per SIMD + and 1920 SPs is true would mean more SIMDs

Cayman will use more SPU and improve utilization over Cypress.
 
Lord_Exodia said:
I guess he's thinking that by increasing the Shader count and not increasing the ROPs you will bottleneck gains in some instances.

Correct me if I'm wrong Snowdog but are you thinking that the extra Shaders will be a waste at times due to a lack of ROPs, and only 256 bit memory interface, and at times it will impact the performance?

It sounds a little logical but as Lorien said we dont know what speed gddr AMD is using and there could be a massive memory bandwidth boost which could open up the bottleneck (if any) a little more, theoretically of course
Click to expand...

If the balance was close to correct in 6870, then it seems that it will be imbalanced when you drastically ramp up SPs but don't touch ROPs/Memory bus.

Weren't you slightly surprised to see the same ROP/Stencil count as 6870?

Certainly there will be some more headroom from Faster ram, faster clock speeds for the ROPs (but 6870 is already running fast). I think the Telegraphs only a modest SP increase.

I suspect Cayman is only going to be about Cypress size. I had speculated this before just because it would need to be limited if they were going to do a Dual GPU card.

After seeing this slide I think it is even more likely that we will be looking at another ~2.1Billion transistor chip with a Modest overall performance bump, that will just be a slight bit ahead of GTX 480. Still quite the feat with a much smaller chip.
 
Snowdog said:
If the balance was close to correct in 6870, then it seems that it will be imbalanced when you drastically ramp up SPs but don't touch ROPs/Memory bus.

Weren't you slightly surprised to see the same ROP/Stencil count as 6870?

Certainly there will be some more headroom from Faster ram, faster clock speeds for the ROPs (but 6870 is already running fast). I think the Telegraphs only a modest SP increase.

I suspect Cayman is only going to be about Cypress size. I had speculated this before just because it would need to be limited if they were going to do a Dual GPU card.

After seeing this slide I think it is even more likely that we will be looking at another ~2.1Billion transistor chip with a Modest overall performance bump, that will just be a slight bit ahead of GTX 480. Still quite the feat with a much smaller chip.
Click to expand...

Perhaps increasing the ROPs would increase the TDP and temperature significantly. If the rumors are to be believed the Cayman is supposed to be hotter than the Barts with a TDP of 255 (if I'm still up to date) then I think they may have had issues increasing the ROPs so they probably increased the Shaders and memory bandwidth all while refining everything.

The end result is if it's 20% faster than GTX 480 which is 10% faster than 5870 that is still a 30 % boost over 5870 and nothing to scoff at. being that it's a refresh. That was with unoptimized drivers I might add.

So I gather with optimized drivers it will be 25% faster than GTX 480 and the GTX 580 may be 30-35% faster and maintain a 5-10% lead just like last time around. Antilles however, that is another story. Nvidia may have an answer with dual GF114 or a brute force power hog dual GF570. I definitely expect a dual card being that nvidia seems to have improved the TDP on the GF11x series of GPU's Time will tell.
 
The entire ROP argument is silly

Look at 3870 -> 4870. They both had 16 ROPs, but did that not improve performance greatly?

Besides, look at some of the Barts OC numbers... Barts benefits more from memory OC than core clock, so the increased core clock does not necessarily translate to more ROP performance.

Either way, given that this is supposed to be either AMD's biggest GPU, or close to biggest GPU, one can already have a feel for the fact that it's going to be a beast.

(Plus, a lot of the SP:ROP ratio was based on 5D, which was inefficient... if they do move to 4D, that changes things big time)
 
ZerazaX said:
Besides, look at some of the Barts OC numbers... Barts benefits more from memory OC than core clock, so the increased core clock does not necessarily translate to more ROP performance.
Click to expand...

Which is an argument that memory is already a bottleneck. My other concern.
 
Status
Not open for further replies.
Back
Top