So we'd have 48*512=24576 GPU-cores - whatever those are, exactly. Each one of those should then achieve 24.576/192.000= 128 GFLOPS?
Last edited by a moderator:
NVIDIA GT200 Rumours & Speculation Thread
- Thread starter Arun
- Start date
- Status
So we'd have 48*512=24576 GPU-cores - whatever those are, exactly. Each one of those should then achieve 24.576/192.000= 128 GFLOPS?
Any source? I think such a big difference is unbelievable but IF true it could be very nice
Well it better be, because otherwise it will be slower then a 9800GX2 most of the times..
Umm, no.
192TFlops / 24576 cores = 7.8GFlops/core. When you include the units, it's harder to go wrong
Or 4ops over 1.95Ghz. double MADD (for 64bit)? Or is the core-count nuts?aaronspink
Veteran
Umm, no.
192TFlops / 24576 cores = 7.8GFlops/core. When you include the units, it's harder to go wrong
um, core count? First of all I can GUARANTEE that the core count is wrong. GUARANTEE! Core count is either 4x10, 4x8 or 4x16.
Aaron Spink
speaking for myself inc.
Not sure if we've had this link yet or not....
http://www.vr-zone.com/articles/GeForce_9900_GTX_&_GTS_Slated_For_July_Launch/5718.html
http://www.vr-zone.com/articles/GeForce_9900_GTX_&_GTS_Slated_For_July_Launch/5718.html
Berek
Regular
What is more important on these cards... a 512-bit memory interface, or GDDR4 or 5? I am not too surprised they are sticking to GDDR3, although GDDR5 could have helped as it is quite a promising memory technology over 3 and especially 4.
This also doesn't state if the cards are DX10 or DX10.1... I would hope DX10.1.
Heh -- I'm with you on this one. But the interesting question, seems to me, is the organization of the ALUs within a core. If there are 512 "somethings" within the 4xGT200 Tesla, and 48 of them yield 192T, something is mighty odd.
You can run a G80 at 4Ghz, so the "something" == "ALU". But if I took a vote on the probability of 4Ghz, I'm betting no one would vote that as likely. We seem likely stuck around 2Ghz, which implies 4flops/cycle, which implies more than just one MADD, or "something" == "double-wide ALU". At anyrate, under that scenario, it is hard to call that an "ALU", and I don't know what to call it.... Math-pipe? Bleah....
Another possibility is that it runs at 2.6Ghz, and nVidia is counting 3flops/cycle.
Or we assume the article is misleading.
Other ideas?
Who knows wth Nvidia is counting. They can't seem to make up their minds. From their Tesla product page:
Overview said:
Specifications said:
Up to 350 Gflops yet 512 Gflops peak? And where did they get 430 Gflops from on a 1.35Ghz shader clock? They're contradicting themselves all over the place.
350 ~= 1.35*128*2
430 ~= 1.35*128*2.5
512 ~= 1.35*128*3
50% is the maximum utilization of the 'Missing MUL' (TM) that has ever been proven on an old driver release from early 2007. I haven't tested that on a recent CUDA release, but I assume this implies that it now exposes the MUL in the same way as that old driver release and that the hardware isn't capable of anything more than that. Of course, that is in an incredibly specific and not-at-all-real-world case.
430 ~= 1.35*128*2.5
512 ~= 1.35*128*3
50% is the maximum utilization of the 'Missing MUL' (TM) that has ever been proven on an old driver release from early 2007. I haven't tested that on a recent CUDA release, but I assume this implies that it now exposes the MUL in the same way as that old driver release and that the hardware isn't capable of anything more than that. Of course, that is in an incredibly specific and not-at-all-real-world case.
My assumption based on the probailities what Nvidia is keen to do for their GT200 :
GT200 : 240SP
Die Size : Slighly bigger than G92 at 55nm
Performance : 2 times as the 8800GTX
512Bit Memory controller
The most underestimate GPU ever.
Credibility : 70%.
Last edited by a moderator:
If you clock Shader clock at four times of the domain clock speed, the latency as well as the branch performance may drop a bit, as if the performance may be not as expected.
Jawed
Legend
I can see it now: Spring 2007 NVidia decides that Crysis is such a monster that the GPU planned for the end of 2007 needs a rethink and so is cancelled in favour of a major improvement. GT200 is that Crysis optimised GPU.
Jawed
leoneazzurro
Regular
It would be a very good GPU, even if I see strong difficulties in seeing this:
GT200 : 240SP
Performance : 2 times as the 8800GTX
512Bit Memory controller
in the same sentence with:
Die Size : Slighly bigger than G92 at 55nm
jimmyjames123
Regular
I'm curious why you say "the most underestimate[d] GPU ever". That implies that it is way faster or better or more feature filled than anyone ever expected it to be, and as far as I know, most people here expect it to be much faster than the 8800 GTX.
Given that 9800GTX was almost exactly the same as 8800GTX as far as performance goes, just more inexpensive, I'm not expecting much from a 9900GTX. How much of a jump can that be. For all we know the 9900GTX is just the G92b, not the G200 (or whatever the new designation will be.)
That said Nvidia seems to have decided on a completely random numbering scheme so for all we know 9800 -> 9900 will be a 100% performance jump while 8800 -> 9800 turned out to be like a 5% jump.
That said Nvidia seems to have decided on a completely random numbering scheme so for all we know 9800 -> 9900 will be a 100% performance jump while 8800 -> 9800 turned out to be like a 5% jump.
Well it would be great to see such a powerful GPU with such a small die but it is not possible imo to do GPU with this specs and with die size as small as ~G92 in 55nm.
Moreover i wonder if what VR-zone have said about Rv770xt performance (they are saying they have got info Rv770Xt is 2x faster than Radeon 2900Pro in 3DMarko06) is true. If yes it could mean AMD has only slightly slower than GT200 (if we take rumours that GT200 has 2xG80/G92 performance) but significantly cheaper in produce (smaller die thanks to 55nm).
Even what Vincent has said is true about GT200 which has 3x performance of GF8800GTX in 3DMarkVantage (a would ask you once again - what rumours have said that? Any reliable source?)
I believe it means that GT200 @ 55nm would be ~same size as G92 (65nm) is
Considering the cost of high end gpu is relatively cost-sensitive for either Nvidia or ATI, I still believe that GT200 is just a refined ASIC, which is similar to the transistion from NV40 to G70.
Even though GT200 is bigger than RV770 at some extent, the die size of GT200 may be approximately twice as big as RV770.
Note : The rumor about Vantage mark score is not as valid as it seems to be.
- Status
