It makes perfect sense. A single G80 is rougly 520 gflops on paper and 430 in real world applications. So that would mean it would need rougly twice the peformance to achive this mark since nvidia claimed it would be *over* 1tflop in performance.Originally Posted by LordEC911
How does that make any sense whatsoever?
A near 1TFlop card doesn't mean it will DO 1TFlop, if they leave it like the G80 it will only be doing some ~600GFlops.
If you go by the logic that die-space is rougly equated with shading units which rougly translates into raw peformance, the G92 would need to be twice as big as their previous G80s in order to achive this goal.
That would mean that in order to produce a chip that had twice as many shading units would by the exact same size as current G80s if it were produced on a 45nm platform. Logic dictates that since they will not make it on a 45nm platform since not even TSMC is capable of this right now. That would mean at 65nm (which will most likely be the process of choice) the "G92" chip would be rougly 20% larger than current G80s on a 90nm platform.
#1 G80s are at the absolute limit for heatsink weight, so the new card *CANNOT* have a higher TDP or they risk losing their PCI-E certification.
#2 G80 yields are not good which is why there are many 96sp 8800GTSs and no cut-down card with the full shading units enabled. In addition the sheer size of the die makes its core extremely expensive to produce at a very mature 90nm process. An unproven 65nm,55nm or even 45nm process would have an even higher defect rate not to mention has the potential to be very leaky which would certianly revoke their PCI-E liscense for that card *IF* they could get one to work properly from the get-go.
Given this information the G92 *cannot* have a larger die than current G80s and in addition cannot have a higher TDP either. That means that the G92 as a single die-single card is *not viable* if it is to go beyond 1tflop.
However, if the G92 was infact a reduced G80 with 512bit ram or even 256bit ram, two of those cards in SLI would easily reach 1tflop and would be much easier to produce since the die would be significantally smaller. Even if the defect rate is high the sheer number of cores would offset this much like Winchester did for AMD.
If I were to take a guess from a buisness standpoint as to what would be the most profitable per peformance for nvidia it would be a 96sp or a 128sp G80 like core with a 256bit memory controller. A card with those specs on a 65nm process would provide a tremendous amount of profit @ a $250 price range and $400-550 as a dual pcb version.
Also the manufacturing cost would be substantially cheaper since their mid and high end use the same GPU and potentially PCB, which is exactally what they did when they released the G71.
Reply With Quote
Bookmarks