Freak'n Big Panda
Regular
Yeah true, it's just that those numbers don't mean anything on their own... well 1/2 the FLOP rate I guess but still. I'm nitpicking.
NVIDIA GT200 Rumours & Speculation Thread
- Thread starter Arun
- Start date
- Status
Yeah true, it's just that those numbers don't mean anything on their own... well 1/2 the FLOP rate I guess but still. I'm nitpicking.
Magnum_Force
Newcomer
Thanks for the explanation guys, it makes alot more sense now 
So RV670 is 10 flops per sahder unit (5 MADD's), there are 16 units in an array, and it has 4 array's.
So basically my calculation only displayed the number of MADD's a sec each card could do, not not the FLOP's (which would basically 2x the final number in my calc, missing the zero's ofc).
So RV670 is 10 flops per sahder unit (5 MADD's), there are 16 units in an array, and it has 4 array's.
So basically my calculation only displayed the number of MADD's a sec each card could do, not not the FLOP's (which would basically 2x the final number in my calc, missing the zero's ofc).
Freak'n Big Panda
Regular
yeah exactly
BS? That was the first leak that correctly hinted at 512-bit 1GHz GDDR3. I'd give them a little bit more credit than that myself!
Arun it states 9800GTX@55nm; one lucky hit doesn't make all the other mistakes right does it?
I only see some forums
Yes, exactly. Sorry it was wrong link.
384SP? If GT200 really has 384 SP (clocked about 1,5Ghz) then there is no way that it won`t be at least 80%-2 times faster than the fastest G92 (i mean GF9800GTX).
I think GT200 HAS TO BE about 2x faster than the fastest G92 because, we should remember, that it is a HIGHEND GPU so it will probably be released to fight against r700 which is supposed to have 2xRv770 performance so if Rv770 is supposed to be about 50-70% faster than Rv670 GT200 should be at least 50-60% faster than Rv770.
Moreover GT200 is a huge die (about G80 size - but maybe the reason is that it is is still 65nm GPU) and Rv770 is even smaller than upcoming G92B then it seems ATI has a big advantage in this case.
I still hope GT200 is next "big thing" (i mean in performance) from NVIDIA like G80 and NV40 were before
384SP? If GT200 really has 384 SP (clocked about 1,5Ghz) then there is no way that it won`t be at least 80%-2 times faster than the fastest G92 (i mean GF9800GTX).
I think GT200 HAS TO BE about 2x faster than the fastest G92 because, we should remember, that it is a HIGHEND GPU so it will probably be released to fight against r700 which is supposed to have 2xRv770 performance so if Rv770 is supposed to be about 50-70% faster than Rv670 GT200 should be at least 50-60% faster than Rv770.
Moreover GT200 is a huge die (about G80 size - but maybe the reason is that it is is still 65nm GPU) and Rv770 is even smaller than upcoming G92B then it seems ATI has a big advantage in this case.
I still hope GT200 is next "big thing" (i mean in performance
) from NVIDIA like G80 and NV40 were before 
IF the rumored ~250mm² are true, then i do not see much of a problem wrt financial competition. True, AMD may still have the advantage of using finer 55nm tech, but otoh they have both transistors not contributing to 3D performance (i.e. 2D, UVD which is supposedly on each chip) and the higher costs of board layout with two GPUs and eventually a still necessary PCIe-Bridge-Chip. Plus, their FB-sizes (though rumor has it, that this has been 'fixed' - well, wasn't really broken, but necessary 'til now) don't double, which is a major disadvantage of both 9800GX2 and HD3870 X2 which are supposed to be 'enthusiast' solutions, but do not have the FB-size of really supporting that badge.
Nvidia otoh may run into trouble with their large chip, as yields will definitely not be so good as with two smaller designs unless they have a massive amount of redundancy and maybe even release GT200 into the wild with not all (maybe 7 out of 8?) functional units enabled. Though more fine grained redundancy would be a better move in my books.
That said, I'm really looking forward to see newer, more powerful GPUs after more than 18 months of stagnation - the curse of markets with small competition.
Disharmonic
Newcomer
Just a small correction. From 780G onwards all AMD chips will no longer have hardware 2D acceleration.
Could you clarify this statement? Has this been packed into the RBEs or is it an SP thing?
If he's saying what I think he's saying then the same has actually been the case since R6xx.
Graphics processors used to have a "VGA" engine for doing 2D and additionally hardware for specific tasks for acceleration of Windows (prior to the Aero desktop). R6xx now translates all 2D calls into 3D calls in the Command Processor and they are executed over the unified 3D pipeline.
Graphics processors used to have a "VGA" engine for doing 2D and additionally hardware for specific tasks for acceleration of Windows (prior to the Aero desktop). R6xx now translates all 2D calls into 3D calls in the Command Processor and they are executed over the unified 3D pipeline.
There's no specific VGA engine, no.
Sounds more than reasonable; graphics units introduced 2Dvia3D in the mobile/PDA market even on non unified HW with the first generation of 3D devices. Removing redundancy is always smart, irrelevant how many times more critical transistor budgets/power consumption are in the small factor markets.
If those supposed 384SPs would be exactly as the G92 SPs at 1.5GHz that would be:
384 * 3 FLOPs * 1.5 GHz = 1.73 TFLOPs
The real question is if that thing truly has as many SPs and if yes how "deep" each exactly will be. What guarantees you that it doesn't have less than 384SPs yet capable of far more FLOPs instead?
As I said before w/o knowing the real internas of a future architecture (if there are any noteworthy changes compared to today's GPUs that is of course), any sort of number-chunking is fairly useless.
By the way if the SPs haven't changed, then squeezing 384 of those into roughly 480mm^2@65nm might be a pretty tough call.
UVD was 4.7mm2 on 65nm. That would make it about 3.5-4mm² on 55nm. Since native 2D engines have been dropped some time ago, I consider this hardly noteworthy. Furthermore, i´m not sure why you come to the conclusion that AMD´s higher PCB costs. Higher than a single RV670 SKU? Of course. Higher costs compared to an upcoming high-end NV SKU? Certainly not the case.
However, the bridge definately should go away, agreed. Which leaves us with the "wasted" framebuffer use, which should also be gone soon. I´m not completely sure how far research has come, regarding the now more widely recognized micro-stutters, but this should also be solvable.
Personally, I think the direction they´re heading is well thought-out. If every new ASIC they produce has RV670-like qualities and gives them at least something that is competitive, AMD is looking at a very bright future in the desktop GPU business. NV already did some pretty unconventional moves, which I didn´t see in years. If AMD can keep their release schedule up, this will get quite interesting.
These are certainly good and valid points, but it´s questionable if NV really wants to disable some units on an high-end ASIC, where the margins are very high to begin with.
On a more general note:
Like Ailuros already wrote, i´m not exactly sure what the general consensus is, regarding "GT200". Personally, if I read numbers like the stated 384 SPs @ 1.5GHz, I´m not sure you can do that just yet (on 65nm/55nm), even on a (fully) custom-designed ASIC. And apart from what Ailuros wrote, the heat output of that ASIC would be enormous when running at full load. However, since currently there is only G92 to compare with, it´s hard to guess something like that. For one, I would be quite surprised if they didn´t change the SPs, that would be very unlike NV. Also, with that many SPs, think about the other units they would need, to keep that thing balanced.
Since I´m not exactly in the market for high-end GPUs, I´m really looking forward to RV770.That said, I'm really looking forward to see newer, more powerful GPUs after more than 18 months of stagnation - the curse of markets with small competition.
Jawed
Legend
Apart from double-precision support, I don't think the functionality will change. Or, if you like, DP is such a significant change that that will be all. You could argue for some improvements in throughput of the new SM4 instructions (e.g. integer divide) but I haven't got any basis for that point of view.
It'll be interesting to see whether NVidia increases the register file size, per multiprocessor, or whether they increase the size of the parallel data cache (of course they could do both).
Still no clue as to whether D3D10.1 support will appear. The work required in TMUs and ROPs is presumably non-trivial.
Jawed
- Status
