(8-17-13) AMD 9970 28nm vs. 20nm Poll

[raghu78]

The 9970 will be a 28nm X2 7850 with hardware frame pacing...to take on 780, I dont think AMD is doing a big die.....

lol :thumbsdown:

I would be irate if this is actually what happens

don't worry. its not going to happen

 

[SolMiester]

I would be irate if this is actually what happens

So, when was the last time AMD did big die?....It makes sense, hardware metering on the respin started 12 mths ago after AMD acknowledged the CF issues...

I know a lot of fanboys are hoping for big die, but its not gonna happen......

 

[3DVagabond]

So, when was the last time AMD did big die?....It makes sense, hardware metering on the respin started 12 mths ago after AMD acknowledged the CF issues...

I know a lot of fanboys are hoping for big die, but its not gonna happen......

Fanboys? How about consumers who want competition to bring down prices?

 

[raghu78]

I know a lot of fanboys are hoping for big die, but its not gonna happen......

this says more about you than others. the irony being you are what you accuse others of. :thumbsup:

 

[Saylick]

So, when was the last time AMD did big die?....It makes sense, hardware metering on the respin started 12 mths ago after AMD acknowledged the CF issues...

I know a lot of fanboys are hoping for big die, but its not gonna happen......

2900XT (R600) was the last time they did a big die, if I'm not mistaken. Everything else beyond that has been smaller than 400 mm^2.

Here's how I see it:

If AMD is to stay relevant for the next couple of months, they will need to either A) lower prices on existing cards, which is not much of a viable option on its own imo, B) release something worthy of challenging nVidia at the high end of the spectrum, OR C) Do both A and B. People on this forum seem to agree that Option B or C is the direction AMD is going to go in the next couple of months. With that said, I don't see much sense for AMD to market a dual GPU card as x970 series; that designation is typically reserved for x990 cards. Secondly, with regards to what 9970 is going to end up being, it can either act as a hopefully short-term stop-gap until 20 nm while matching 780's performance at 28 nm and at a lower cost, OR if the folks at TSMC somehow got their act together, AMD could pull out a sub-400 mm^2 20 nm next-gen beast.

The question that remains is which outcome is more likely; would it be A) short-term 28 nm GTX 780 contender, or B) back-to-basics small-die strategy 20 nm? If the truth is A, then I believe we would almost certainly see a 400 mm^2+ die. There is simply too much of a performance gap between Tahiti and the 780 for AMD to close without using a larger die and/or significant power/area efficiency enhancements. If the truth is B, then there is not much we can discuss besides asking how the hell did AMD, let alone TSMC, manage to pull out a 20 nm class card this early... unless they are simply doing a paper launch this year with a hard launch sometime next year.

 

[el etro]

Dreaming bigger this time: A ~550mm² GPU to outperform everything we have today.

 

[SolMiester]

this says more about you than others. the irony being you are what you accuse others of. :thumbsup:

Not half as much as you......

 

[SolMiester]

Fanboys? How about consumers who want competition to bring down prices?

Fanboys dont want that?...and not sure how consumers wanting a product equals = capable.....At any rate, 2 x small die does = competition for big die....

 

[raghu78]

Not half as much as you......

is it. you come in here calling people who want to see good competition as fanboys. its clear you want to keep worshipping Nvidia. :thumbsdown:

 

[Saylick]

FFS guys... Keep the personal talk out of the thread.

Fanboys dont want that?...and not sure how consumers wanting a product equals = capable.....At any rate, 2 x small die does = competition for big die....

If you want 7850 X2, you could get that TODAY by buying 2 7850s. I see little to no reason why AMD would want to release a product which is essentially 2 Pitcairn GPUs on the same board especially since it would be potentially harmful for their public image. Doing something like that falls in the same category of badge renaming, imo.

If what I hear is true, i.e. AMD is indeed making big enough splashes for us to look forward to the 9xxx series, I believe they will have something new for us, and by "new" I mean a new die.

 

[RussianSensation]

The 9970 will be a 28nm X2 7850 with hardware frame pacing...to take on 780, I dont think AMD is doing a big die.....

This theory makes no sense. AMD already has HD7990 selling for $700. What is the point of a 7850X2 card at $550? Firstly, people looking at single GPU performance won't touch it with a 30 foot pole. Secondly, 1Ghz 7970 is going for $290, which means 2 of those are just $580. Why would anyone pay $500+ for 7850X2 card?

Making HD7870X2 card also makes no sense since 7870s are dipping into $170-200 range. AMD won't release such a card for only $350.

IMO, what AMD needs to do is release a fused HD7870 x2 on 1 die and forget about double precision until they can afford to bring it back on 20nm. Such a chip would beat the Titan and come in at 425mm2 or so and still manage just 250W of real world power consumption or less.

 

[raghu78]

This theory makes no sense. AMD already has HD7990 selling for $700. What is the point of a 7850X2 card at $550? Firstly, people looking at single GPU performance won't touch it with a 30 foot pole. Secondly, 1Ghz 7970 is going for $290, which means 2 of those are just $580. Why would anyone pay $500+ for 7850X2 card?

RS do you even what to reply to that crap. you know what that person is doing right ?

IMO, what AMD needs to do is release a fused HD7870 x2 on 1 die and forget about double precision until they can afford to bring it back on 20nm. Such a chip would beat the Titan without a problems and come in at 425mm2 or so and still manage just 250W of real world power consumption or less.

not probable. AMD is committed to HPC and gaming. so the flagship Volcanic Islands GPU is a compute part with double precision. imo it is a compute/gaming chip with a die size of 450+ sq mm. I see atleast 2 salvage chips. so they will fill USD 400 - 600 with 3 SKUs.

I hope they refresh Pitcairn with 1920 or 2048 sp and a 256 bit memory bus at 7 Ghz and price it at USD 300. should be below 300 sq mm. much more profitable for AMD than selling 365 sq mm Tahiti with 3GB VRAM .

 

[Saylick]

IMO, what AMD needs to do is release a fused HD7870 x2 on 1 die and forget about double precision until they can afford to bring it back on 20nm. Such a chip would beat the Titan without a problems and come in at 425mm2 or so and still manage just 250W of real world power consumption or less.

I'm keeping my fingers crossed for 3x Bonaire.

2688 SP
168 Texture Units
48 ROPs
384-bit Bus Width
3GB GDDR5
~250W
... and hopefully not 2 but 3 ACEs along with improvements to the ROPs.

 

[raghu78]

I'm keeping my fingers crossed for 3x Bonaire.

2688 SP
168 Texture Units
48 ROPs
384-bit Bus Width
3GB GDDR5
~250W
... and hopefully not 2 but 3 ACEs along with improvements to the ROPs.

its pretty much a given that AMD is increasing the front end (ACE , geometry engine, rasterizer) and the back end (ROPs). 3 to 4 (ACE, geometry engines and rasterizer). 48 ROPs.

2816 sp , 44 CU, 4 ACE, 4 geometry engines, 4 rasterizer, 48 ROPs is my guess.

 

[Saylick]

its pretty much a given that AMD is increasing the front end (ACE , geometry engine, rasterizer) and the back end (ROPs). 3 to 4 (ACE, geometry engines and rasterizer). 48 ROPs.

2816 sp , 44 CU, 4 ACE, 4 geometry engines, 4 rasterizer, 48 ROPs is my guess.

Are the number of CUs tied in any way to the number of ROPs?

Tahiti is 2048 SP with 32 ROPs for a SP:ROP ratio of 64
Pitcairn is 1280 SP with 32 ROPs for a SP:ROP ratio of 40
Bonaire is 896 SP with 16 ROPs for a SP:ROP ratio of 56.

What you have proposed has a SP:ROP ratio of 58 and 2/3rds.

 

[raghu78]

Are the number of CUs tied in any way to the number of ROPs?

Tahiti is 2048 SP with 32 ROPs for a SP:ROP ratio of 64
Pitcairn is 1280 SP with 32 ROPs for a SP:ROP ratio of 40
Bonaire is 896 SP with 16 ROPs for a SP:ROP ratio of 56.

What you have proposed has a SP:ROP ratio of 58 and 2/3rds.

no ROPs are linked to the memory controller and have nothing to do with sp. 64 sp make 1 compute unit or CU. these CUs are generally combined in groups of 4 and sometimes 3. so Tahiti has (4 CU x 4) x 2 = 32. Pitcairn has (4CU, 3 CU, 3 CU ) x 2 = 20 CU. Bonaire has (4CU, 3CU) x 2 = 14 CU.

http://www.techpowerup.com/reviews/AMD/HD_7970/images/architecture.jpg

http://www.techpowerup.com/reviews/AMD/HD_7850_HD_7870/images/slide2.jpg

http://www.techpowerup.com/reviews/Sapphire/HD_7790_Dual-X/images/intro3.jpg

so 2816 sp with 4 ACE, 4 geometry engines, 4 rasterizer would be built as ( 4 CU, 4CU, 3CU) x 4 = 44 CU. If Hawaii were 3072 sp then it would be (4CU x 3) x 4 = 48 CU.

 

[Saylick]

no ROPs are linked to the memory controller and have nothing to do with sp. 64 sp make 1 compute unit or CU. these CUs are generally combined in groups of 4 and sometimes 3. so Tahiti has (4 CU x 4) x 2 = 32. Pitcairn has (4CU, 3 CU, 3 CU ) x 2 = 20 CU. Bonaire has (4CU, 3CU) x 2 = 14 CU.

http://www.techpowerup.com/reviews/AMD/HD_7970/images/architecture.jpg

http://www.techpowerup.com/reviews/AMD/HD_7850_HD_7870/images/slide2.jpg

http://www.techpowerup.com/reviews/Sapphire/HD_7790_Dual-X/images/intro3.jpg

so 2816 sp with 4 ACE, 4 geometry engines, 4 rasterizer would be built as
( 4 CU, 4CU, 3CU) x 4 = 44 CU. If Hawaii were 3072 sp then it would be
(4CU x 3) x 4 = 48 CU.

I understand the ROPs being tied to the memory controller, but there has to be a reason why the SP:ROP ratio thus far has been a round number, i.e. integer, excluding cut-down dies of course. :hmm:

 

[raghu78]

I understand the ROPs being tied to the memory controller, but there has to be a reason why the SP:ROP ratio thus far has been a round number, i.e. integer. :hmm:

if there is such a relation between ROP and sp having an integer ratio then then 2688 (44 CU) or 3072 sp(48 CU) are the only possible combinations. 2688/ 48 = 56 , 3072/48 = 64 . out of these 2 configs, the fact that 3072 sp allows for 3 complete 4CU groups and has symmetricity means its the more probable choice. now that would be a beast chip

 

[3DVagabond]

Fanboys dont want that?...and not sure how consumers wanting a product equals = capable.....At any rate, 2 x small die does = competition for big die....

How does someone wanting AMD to come out with a single GPU competitor to GK110 mean they are fanboys?

It was pretty obvious when nVidia released Titan and it outsold the GTX-690 overall in just a few months that 2x smaller GPU's is not equal/competition for a large single GPU. Even when the smaller duallie outperforms the single.

I've thought about where this dual gpu rumor could have gotten started, assuming there's often some truth to rumors. The only scenario that makes sense to me is if AMD is going to graph 2 smaller gpu's together in some sort of modular construction. I'm not sure if this is practical with GPU's, or even possible. If anyone has the tech to do it though, it would be AMD, if you consider they do it with CPU's already.

Edit: Hadn't read where RS posted a similar idea with 2xPitcairn on one die before I responded. I think we are both just purely speculating, though. I know I am, anyway.

 

[Saylick]

if there is such a relation between ROP and sp having an integer ratio then then 2688 (44 CU) or 3072 sp(48 CU) are the only possible combinations. 2688/ 48 = 56 , 3072/48 = 64 . out of these 2 configs, the fact that 3072 sp allows for 3 complete 4CU groups and has symmetricity means its the more probable choice. now that would be a beast chip

Hm... Well, if the extrapolation works, 3072 SP might be within the ballpark for 780-esque, and maybe even Titan-esque, performance...

680/770 (1536 SP) vs. 7970 (2048 SP)
780 (2304 SP) vs 9970 (3072 SP?)

 

[raghu78]

Hm... Well, if the extrapolation works, 3072 SP might be within the ballpark for 780-esque, and maybe even Titan-esque, performance...

680/770 (1536 SP) vs. 7970 (2048 SP)
780 (2304 SP) vs 9970 (3072 SP?)

actually Tahiti is unbalanced and stands to gain perf/sp by doubling front end and increasing ROPs to 48. so a Hawaii at 2048 sp could be 10% faster clock for clock. at 975 mhz it would still be 5% faster.

Throw in 50% more sp and you are looking at close to 40 - 45% more performance. so if anything a 3072 sp will edge out a Titan and match a 2880 sp GK110.

 

[Vesku]

not probable. AMD is committed to HPC and gaming. so the flagship Volcanic Islands GPU is a compute part with double precision. imo it is a compute/gaming chip with a die size of 450+ sq mm. I see atleast 2 salvage chips. so they will fill USD 400 - 600 with 3 SKUs.

I hope they refresh Pitcairn with 1920 or 2048 sp and a 256 bit memory bus at 7 Ghz and price it at USD 300. should be below 300 sq mm. much more profitable for AMD than selling 365 sq mm Tahiti with 3GB VRAM .

I've already said it, people discounting a "big" die 28nm chip are ignoring that HPC is where the higher margins are. AMD has even started merging APUs into their FirePro branding, although with limited success. Combine desires for the HPC market and the probability that 20nm transistor costs will not be much improved over 28nm if at all and it looks quite presumptuous to decide AMD wouldn't at least have considered a larger die size bump than 5870->6970.

 

[SolMiester]

How does someone wanting AMD to come out with a single GPU competitor to GK110 mean they are fanboys?

It was pretty obvious when nVidia released Titan and it outsold the GTX-690 overall in just a few months that 2x smaller GPU's is not equal/competition for a large single GPU. Even when the smaller duallie outperforms the single.

I've thought about where this dual gpu rumor could have gotten started, assuming there's often some truth to rumors. The only scenario that makes sense to me is if AMD is going to graph 2 smaller gpu's together in some sort of modular construction. I'm not sure if this is practical with GPU's, or even possible. If anyone has the tech to do it though, it would be AMD, if you consider they do it with CPU's already.

Edit: Hadn't read where RS posted a similar idea with 2xPitcairn on one die before I responded. I think we are both just purely speculating, though. I know I am, anyway.

3d, you are just twisting what I said...i dont know why you are saying that...i said nowt about competition, you did?...

 

[SolMiester]

This theory makes no sense. AMD already has HD7990 selling for $700. What is the point of a 7850X2 card at $550? Firstly, people looking at single GPU performance won't touch it with a 30 foot pole. Secondly, 1Ghz 7970 is going for $290, which means 2 of those are just $580. Why would anyone pay $500+ for 7850X2 card?

Making HD7870X2 card also makes no sense since 7870s are dipping into $170-200 range. AMD won't release such a card for only $350.

IMO, what AMD needs to do is release a fused HD7870 x2 on 1 die and forget about double precision until they can afford to bring it back on 20nm. Such a chip would beat the Titan and come in at 425mm2 or so and still manage just 250W of real world power consumption or less.

The 7990 doesnt have frame pacing, only a driver that still doesnt work as well as NV hardware frame pacing...thats what I believe the respin is all about....

Jesus, say anything not in line with the knob ends in here and you are called out.....reading comprehension in here goes awol half the time....I never said people wanting competition were fanboys and yet.....?

 

[Shmee]

No accusations of fanboism please. Keep this thread on topic.