Late 6970?

[AnandThenMan]

I dont see why they would be utilizing a sub 100 dollar chip to do it.

Do you realize that Opteron will eventually be an APU type product? I don't understand why you only think the APU is confined to the lowest end.

Intel is going to go down the same path, Larrabee was an over ambition and flawed attempt at it. But some version of Fusion, Larrabee etc. will become the common processor. The trends are obvious.

 

[WelshBloke]

Yeah I can see APUs stealing even more of the laptop/notebook market share... instead of haveing 61% amd might raise to like ~80%+. It might even effect the low end desktops abit, so they take abit of market share on that front too.

The apu wont have any effect on the mid-high, as I see it.

Depending on the power of the on board video it might even drive sales for discrete cards on the desktop as people get a taste for PC gaming again.

 

[Genx87]

Your thinking seems to be very tightly confined to what the structure of the current market is, you need to think about how the market is evolving.

And the examples cited have everything to do with this.

Please tell me how professionals who demand the highest performance will be happy with an APU designed for the sub 100 market? Once you tell me why they would settle for less performance. Then maybe you can start changing my mind.

 

[Arkadrel]

Welsh Bloke,

the llnao is probably gonna be like a 450 card or something... like was said earlier on in the thread, its not really a top end gameing card ^-^. It ll probably play WoW if thats your thing though.

 

[Genx87]

Do you realize that Opteron will eventually be an APU type product? I don't understand why you only think the APU is confined to the lowest end.

Intel is going to go down the same path, Larrabee was an over ambition and flawed attempt at it. But some version of Fusion, Larrabee etc. will become the common processor. The trends are obvious.

Will the operton provide the same performance as the highest end GPU? I highly doubt it.

 

[Makaveli]

Amd are in that market, they just aren't that successful in it.

I think they will concentrate on pinning down the laptop market first with fusion.

I was thinking along those lines, that is why I said after a few generations of fusion products are behind us.

But it seems from gen post that the drivers are still a major stumbling block for AMD. And my assumption is most businesses won't be as forgiving with broken stuff in the drivers than a consumer would be.

I was looking forward to upgrading my 4890 to the 6970 but with the expected price rape that will ensue. It will probably be well above MSRP for a few months hopefully not. So looks like no new card for me until Q1 at the earliest 2011.

 

[Aristotelian]

Dude do you ever stop I dropped it and you keep going.

My original post was for T2k then you decided to put in your 2c which I never asked for.

So then I responded to you.

NOW FUCKEN DROP IT.

Do not post on a public message board if you do not want answers, it is honestly that simple. I do not need your permission to post on something you wrote, and it sounds like this is news to you. Also, do not presume to tell me when to drop something. My last post conclusively proved you were wrong. Repeating the same thing (that you were talking to T2K) does not change anything. Look at the post sequence. Also, you are not 'dropping it' by responding to my facts with your lies, and then demanding that I drop the subject. You demand nothing here.

Amd are in that market, they just aren't that successful in it.

I think they will concentrate on pinning down the laptop market first with fusion.

Careful chief, he didn't ask you for your '2c', his post was explicitly addressed to Genx87. You may incur his wrath too.

Joking aside, I think it's pretty obvious that what you say is correct. The big piece of the graphics market occurs outside of discrete cards, correct? Intel, if I recall correctly, has the largest share of the graphics market. So AMD is going to want a bit of that pie too.

I think some posts in this thread have evolved into a discussion about fusion chips (cpu and gpu functions together) and discrete gpus. I don't think we'll be seeing such fusion chips overtaking discrete cards for a very, very long time. Has Intel or AMD suggested this would happen eventually?

 

[Arkadrel]

You have to remember at some point, IBM will probably step up their game.

They live off that server market and probably wont allow a challenge by GPGPUs to go them by. Like that super computer for china thingy... thats stepping on IMBs toes I think (they make all those super computers normally).

 

[WelshBloke]

Welsh Bloke,

the llnao is probably gonna be like a 450 card or something... like was said earlier on in the thread, its not really a top end gameing card ^-^. It ll probably play WoW if thats your thing though.

Thats my point though, It will play games, just not very well. Therefore its reintroducing people to the idea of PC gaming and encouraging them to buy better cards.

Thats my plan any way :ninja:

 

[AnandThenMan]

Will the operton provide the same performance as the highest end GPU? I highly doubt it.

Likely much, much higher performance/watt. Much lower latency, vastly higher density, lower power through integration. Lower costs, ability to run general purpose code. And don't forget about shared memory, how much memory can you really stuff into a discreet card? Direct access to a very large continuous memory has many advantages.

It will never provide the ultimate performance, 1 APU vs. 1 discreet card. But consider, using an APU teamed with a discreet card, all from the same vendor. Now you are talking even higher potential.

 

[blastingcap]

HD6970/6950 is NOT is not delayed. The translation is rough, ut I think nApoleon us saying Nov. 22 is the presentation date (to AIBs or the media, I am not sure which). So Nov. 22 was never the launch date to begin with, apparently. Seems like the launch date was and still is Nov. 29, based on the rumor mill. Additionally, Chiphell rumor is that HD6950 is the one with 1536 shaders, with HD6970 having 1920 shaders. That translates to almost 50% more shader power, which makes me wonder how they could get enough memory bandwidth to feed the shaders--unless they go over 256-bit bus, but if they do go over, and 2GB VRAM is accurate, then that implies a 512-bit bus which AMD has shied away from for years now due to cost.

http://translate.google.com/transla...ttp://www.chiphell.com/thread-137198-1-1.html

Additionally, Neliz at B3D was very accurate with Barts forecasts, right down to the SEP of $179/239. (But like everyone else had predicted 1GB VRAM for 6970, perhaps because of that leaked slide last month showing 1GB VRAM as standard for Cayman.)

http://forum.beyond3d.com/showpost.php?p=1492259&postcount=4720

I believe Chiphell and Neliz more than FUDzilla and irresponsible sites that parrot FUDzilla. With the new info on shaders, if it's accurate, I'm going to revise my guesstimate to:

a) HD6970 ties or edges out GTX580
b) at $450 price point
c) with >90% chance of launching prior to Dec. 1, 2010

 

[Arkadrel]

Anandthenman,

Do you know for a fact that the new apus will support hybrid crossfire?

 

[AnandThenMan]

Anandthenman,

Do you know for a fact that the new apus will support hybrid crossfire?

Yes. AMD stated so during this recent analysts day.

 

[AnandThenMan]

I believe Chiphell and Neliz more than FUDzilla and irresponsible sites that parrot FUDzilla.

FUD has the worst record I think I've ever seen on any site. The articles they wrote last year about Fermi were so ridiculous and just flat out WRONG. And they write nothing but double speak and gibberish.

Zero reason to trust them as a source.

 

[Arkadrel]

that is pretty cool indeed.

So Llano + Radeon HD 5670 and maybe ~50-60% performance increase that way?
(the older 4200 IGP + hybrid crossfire didnt really scale that well, 50% was like avg)

 

[Aristotelian]

HD6970/6950 is NOT is not delayed. The translation is rough, ut I think nApoleon us saying Nov. 22 is the presentation date (to AIBs or the media, I am not sure which). So Nov. 22 was never the launch date to begin with, apparently. Seems like the launch date was and still is Nov. 29, based on the rumor mill. Additionally, Chiphell rumor is that HD6950 is the one with 1536 shaders, with HD6970 having 1920 shaders. That translates to almost 50% more shader power, which makes me wonder how they could get enough memory bandwidth to feed the shaders--unless they go over 256-bit bus, but if they do go over, and 2GB VRAM is accurate, then that implies a 512-bit bus which AMD has shied away from for years now due to cost.

If we never had a conclusive launch date to begin with I'm still not sure how anyone could make the claim that it's late.

Thanks for that post. Can those in the know speculate based on the shader count what sort of performance the 6970 might be capable of? I think it was pretty much certain in another thread that the 2GB count was accurate, remember? (It evolved into a '2GB is too much' debate).

 

[happy medium]

That translates to almost 50% more shader power,

Then what? a 5870? I though it had 1600? or the 6850? 960 shaders.
I guess it depend on what kind of shaders they use right? 4d, 5d ect. ect.

 

[busydude]

unless they go over 256-bit bus, but if they do go over, and 2GB VRAM is accurate, then that implies a 512-bit bus

2GB VRAM is possible with 256-bit bus.

 

[AnandThenMan]

that is pretty cool indeed.

So Llano + Radeon HD 5670 and maybe ~50-60% performance increase that way?
(the older 4200 IGP + hybrid crossfire didnt really scale that well, 50% was like avg)

I sure hope so. The way I interpreted what was being said is, AMD is going to the APU eventually across the board, there will be no such thing as just a plain old CPU anymore. So why waste what is onboard if you buy a discreet card.

 

[blastingcap]

Then what? a 5870? I though it had 1600? or the 6850? 960 shaders.
I guess it depend on what kind of shaders they use right? 4d, 5d ect. ect.

Rumor was that 4D clusters could do ~98% as much work as 5D clusters. 1920/4 = 480, and 1600/5 = 320. .98*1.5 = 1.47. Things usually don't scale perfectly and the bottleneck might be somewhere else, or the shaders might not be 1920 (that is just a rumor), hence my conservative estimate for 6970's speed. I'm definitely not saying that 6970 is 47% faster than Cypress XT although that's possibly in theory.

A lot depends on the tessellator, too. Even with all those shaders, 6970 may end up being hamstrung by its tessellator in certain games.

@busy, yeah but 256-bit bus even with 6Ghz RAM means 6970 would only have 25% more memory bandwidth than Cypress XT, so something's gotta give. Either the 1920 shader number is wrong, or 98% number, or the 2GB VRAM, or 256-bit bus, either that or AMD worked some sort of miracle in rebalancing the card or else the card is unbalanced as hell. Keep in mind that Barts seems to be memory-bottlenecked so that last possibility is real.

I smell a price war in the high-end segment in a couple of weeks. Consumers win again. :thumbsup:

 

[busydude]

@busy, yeah but 256-bit bus even with 6Ghz RAM means 6970 would only have 25% more memory bandwidth than Cypress XT, so something's gotta give. Either the 1920 shader number is wrong, or 98% number, or the 2GB VRAM, or 256-bit bus, either that or AMD worked some sort of miracle in rebalancing the card or else the card is unbalanced as hell. Keep in mind that Barts seems to be memory-bottlenecked so that last possibility is real.

Cypress was never memory bottle necked, its shaders were not efficient and did not scale perfectly. Barts, OTOH, is bottlenecked.

ABT's 5770 bottleneck investigation:

The results actually make sense given a 5770 is basically a 5870 chopped in half, and the 5870 was proven to be primarily limited by its core too. If you shrink the processing power and bandwidth in proportion to each other like the 5770 does (i.e. cutting both in half from the 5870), you’ll end up with very similar performance characteristics.

I think AMD, with better memory controller, can easily tackle this problem or lack thereof.

 

[Skurge]

Rumor was that 4D clusters could do ~98% as much work as 5D clusters. 1920/4 = 480, and 1600/5 = 320. .98*1.5 = 1.47. Things usually don't scale perfectly and the bottleneck might be somewhere else, or the shaders might not be 1920 (that is just a rumor), hence my conservative estimate for 6970's speed. I'm definitely not saying that 6970 is 47% faster than Cypress XT although that's possibly in theory.

A lot depends on the tessellator, too. Even with all those shaders, 6970 may end up being hamstrung by its tessellator in certain games.

@busy, yeah but 256-bit bus even with 6Ghz RAM means 6970 would only have 25% more memory bandwidth than Cypress XT, so something's gotta give. Either the 1920 shader number is wrong, or 98% number, or the 2GB VRAM, or 256-bit bus, either that or AMD worked some sort of miracle in rebalancing the card or else the card is unbalanced as hell. Keep in mind that Barts seems to be memory-bottlenecked so that last possibility is real.

I smell a price war in the high-end segment in a couple of weeks. Consumers win again. :thumbsup:

Or its going to use 7Ghz ram, 6Ghz for 6950 and maybe something like 6.5 or 6.8Ghz for the 6970... Im just saying.

7gbps GDDR5 has been in production for months now.

 

[tannat]

Rumor was that 4D clusters could do ~98% as much work as 5D clusters. 1920/4 = 480, and 1600/5 = 320. .98*1.5 = 1.47. Things usually don't scale perfectly and the bottleneck might be somewhere else, or the shaders might not be 1920 (that is just a rumor), hence my conservative estimate for 6970's speed. I'm definitely not saying that 6970 is 47% faster than Cypress XT although that's possibly in theory.

A lot depends on the tessellator, too. Even with all those shaders, 6970 may end up being hamstrung by its tessellator in certain games.

@busy, yeah but 256-bit bus even with 6Ghz RAM means 6970 would only have 25% more memory bandwidth than Cypress XT, so something's gotta give. Either the 1920 shader number is wrong, or 98% number, or the 2GB VRAM, or 256-bit bus, either that or AMD worked some sort of miracle in rebalancing the card or else the card is unbalanced as hell. Keep in mind that Barts seems to be memory-bottlenecked so that last possibility is real.

I smell a price war in the high-end segment in a couple of weeks. Consumers win again. :thumbsup:

Probably better to compare with 6870 than 5870 since cayman ought to be closer to Bart architecture:

1920/1120*0,98*5/4 = 2,1

But this is only a number sofar...
Huge though

 

[busydude]

Probably better to compare with 6870 than 5870 since cayman ought to be closer to Bart architecture:

No.. infact its the opposite. Cayman is supposedly new architecture.. and Barts is a refined er.. tweaked cypress.

 

[tannat]

No.. infact its the opposite. Cayman is supposedly new architecture.. and Barts is a refined er.. tweaked cypress.

No:

Cypress->Bart->Cayman

Bart was supposed to be cypress shaders with cayman architecture

You need to tweak/refine quite much to keep 95% performance with a 25% smaller chip. (Barts versus Cypress) It's quite an astonishing achievement at the same node.