Anyone actually know how much it costs in Watts or area? Can it be very much when Juniper is larger than half a Cypress?
AMD: R9xx Speculation
- Thread starter Lukfi
- Start date
Anyone actually know how much it costs in Watts or area? Can it be very much when Juniper is larger than half a Cypress?
transistor-wise
Juniper - Cypress: 2,07x (exactly half)
RV730 - RV770: 1,85x (RV730 was more than half of RV770)
RV635 - RV670: 1,76x (RV635 was less than half of RV670)
RV410 - R420: 1,33x (RV410 was exactly half of R420)
maybe the R4xx architecture is too old and its non-scalable blocks consumes too much die-size compared to current GPUs, but R6xx and R7xx should be quite close... given the R6xx and R7xx numbers I'd expect that Cypress should be about 1,8-1,85x bigger than Juniper... PCIe, triangle-setup, tesselator, UVD, hub... that should be about the same for both GPUs. But Cypress is 2,07-times bigger though. I think 200-250 milions of transistors in Cypress can be related to GPGPU stuff (or anything what isn't present in Juniper)
Juniper - Cypress: 2,07x (exactly half)
RV730 - RV770: 1,85x (RV730 was more than half of RV770)
RV635 - RV670: 1,76x (RV635 was less than half of RV670)
RV410 - R420: 1,33x (RV410 was exactly half of R420)
maybe the R4xx architecture is too old and its non-scalable blocks consumes too much die-size compared to current GPUs, but R6xx and R7xx should be quite close... given the R6xx and R7xx numbers I'd expect that Cypress should be about 1,8-1,85x bigger than Juniper... PCIe, triangle-setup, tesselator, UVD, hub... that should be about the same for both GPUs. But Cypress is 2,07-times bigger though. I think 200-250 milions of transistors in Cypress can be related to GPGPU stuff (or anything what isn't present in Juniper)
Ok you got a point
But will AMD let Nvidia go alone for the HPC market? It's not just because those Tesla/Kepler cards won't be AMDs, but also because those cards are going to bite into AMD's server market.. i'm just guessing that more Tesla are sold, less Opteron are sold.
DavidGraham
Veteran
I think (And please somebody correct me if I am wrong ) this is a special case , if AMD is capable of doing the same in every iteration of GPU family , then the transition from RV770 to Cypress (smaller process) should have came with minimal increase in die area , After all , the jump from RV670 to RV770 (2.5X shaders & TU ) is bigger than the jump from RV770 to Cypress (2X shaders & TU) , with the exception of DX11 logic of course .
Of course such "miracles" can't happen every generation, but if they're making major changes to shader architecture, why is it impossible for it to bring at least close to similar die-size improvements per unit as RV670>770 brought?
There was something about ATI's tools that kept the density of the SIMDs lower then they should have been for RV670.
The tools for RV770 corrected this and allowed density to reach a point where they should have been for earlier designs.
Unless we think there is something wrong with the density numbers for Cypress, there may not be a single factor that is artificially lowering density that can be removed to such great effect.
The tools for RV770 corrected this and allowed density to reach a point where they should have been for earlier designs.
Unless we think there is something wrong with the density numbers for Cypress, there may not be a single factor that is artificially lowering density that can be removed to such great effect.
Not much actual info, just somewhat official mention of the 6000 series: http://www.youtube.com/user/Thermalrightonutube#p/a/u/0/n46_N939di8
Ok you got a point
But will AMD let Nvidia go alone for the HPC market? It's not just because those Tesla/Kepler cards won't be AMDs, but also because those cards are going to bite into AMD's server market.. i'm just guessing that more Tesla are sold, less Opteron are sold.
Tesla's competition isn't with Opteron.
Why would AMD release a 6870 part that only has 90% of the 5870's performance?
It does not make naming sense, but maybe it does make marketing sense. A 6950 at 350 euros while it should have been a 6850 at less than 350 euros?
Give me a 6850 with 20% over 5870 for 300 euros and we are talking.
Does the freaking NUMBER of the product decide for you if you buy it or not? You're making it sound like you wouldn't buy it if it was named 6180, because you'd feel you're buying an inferior product?
Gee, would you like me to remind you how nv managed to sell G92 all these years?
would they've sold more or less if they branded it 255 or 245?
if it's an 68x0 .. it probable means it's faster than a 67x0. and that's it, cross generation naming has been awkward for ages.
Not that awkward. I mean, clearly, HD 3870 < HD 4870 < HD 5870, and the same holds true for most SKUs.
I guess the HD 5450 is the exception, but no one really cares about raw performance on those cards anyway.
Silent_Buddha
Legend
Yes, but we've never had a situation where the x8xx product was across the board slower than the previous generation x8xx product.
Everyone understands that x8xx should be faster than x7xx. But everyone expects a new generation x8xx product should be faster or at least no slower than last generation x8xx product. The last time there might have been confusion is when we went from HD 2900 XT/Pro to 3870/50. But then again, that's when AMD implemented their new naming scheme, so things aren't exactly comparable there.
So, if 6870 were actually slower than 5870, that would be a slap in the consumers face and make a farce of AMD's own naming scheme.
I'd be quite happy to lambast them the same as I've done to Nvidia if that were the case. I can only hope AMD aren't stupid or desperate enough to do that. Then again they are in the process of getting rid of the ATI brand in favor of the much much weaker AMD brand, so I guess anything is possible.
Either way, I have a hard time giving that chart any validity. Unless as no-X was suggesting they just got their wording wrong and they actually 90% MORE performance rather than 90% of the performance.
But then that still brings into question its validity as I have a hard time believing 6870 is going to be 190% of the speed of 5870.
Regards,
SB
No it doesn't. I just said it doesn't make naming sense.
My last sentence "Give me a 6850 with 20% over 5870 for 300 euros and we are talking" should be more like "Give me a XXXX with 20% over 5870 for 300 euros and we are talking".
My main complaint would focus on the fact that the supposed 6950, which is meant to replace the 5850, would have a 100 euros higher price than the 5850 when it initially launched.
So we would be facing a situation where the 5850 cost 100 euros more than the 4850 it replaced, while the 6950 (or whatever) would cost another 100 euros more than the 5850 it replaced. So where does it stop really?
All that is hypothetical of course. Just commenting on the suggested prices, is all.
Silent_Buddha: Maybe they decided to change the naming scheme again. If you take in account, that they decided to release 4xx mm^2 parts again, it would make sens... It could be a bit similar to X1xxx naming scheme
*6** for mainstream
*7** for mid-range
*8** for high-end
*9** for enthusiast
**5* for cheaper model
**7* for full-speed model
**9* for dual-GPU
HD6870 would be full Barts
HD6970 would be full Cayman
HD6990 would be X2 Cayman
anyway, this is only speculation
*6** for mainstream
*7** for mid-range
*8** for high-end
*9** for enthusiast
**5* for cheaper model
**7* for full-speed model
**9* for dual-GPU
HD6870 would be full Barts
HD6970 would be full Cayman
HD6990 would be X2 Cayman
anyway, this is only speculation
UniversalTruth
Veteran
I don't think ATi spend that much R&D on R800 -- mostly a DX11-upgraded 2*RV770 single ASIC implementation.
Surely something much more grandiose is in the workshop behind the curtain.
Post 5 in this almost one-year-old thread. I think fellix has the same positive feeling.
Similar threads
