NVIDIA GF100 & Friends speculation

The whispers are that there won't be any clocks or performance numbers in this architecture revelation. It'll be great to know what little secrets Nvidia has in store but the wait still won't be over.
 
Well, if you know the internals and account for some other stuff you can probably get within 5% +-.

Without a peformance number and the specific clocks it is not possible to guess the "real" performance of GF100.
And you can't compare GF100 to GT200(b) and claim, that GF100 must be xxx% faster because it has xx(x)% more units. Look at Cypress: Double ALUs, TMUs, ROPs and only 50% faster than the 4890 in <=D3D10.x games.
 
Last edited by a moderator:
Sontin said:
Without a peformance number and the specific clocks it is not possible to guess the "real" performance of GF100.
Maybe not for you.... are you seriously trying to speculate on the concept of speculation? :LOL:
 
We need more data! All this waiting to call Fermi an abject failure is grating on my nerves! LOL JK.

Seriously where are the leaks?!?
 
Seriously where are the leaks?!?

Fermi Fun Fact: Launch products will have less TMU's than GT200.

edit: edited for clarity! (after this post I'm going to know much flak I can take for trusting someone)
 
Last edited by a moderator:
Fermi Fun Fact: Launch products will have less TMU's than GT200.
Less TMUs? Are you serious? GT200 had 80 TMUs. Since Fermi has 8 TMUs per cluster just as previous products, that'll mean only 8 or 9 clusters out of 16 would be enabled (for 64 and 72 tmus, respectively). I could believe maybe one card with only 10 clusters, but less than that???
 
On the triangle setup thing, I think the GF100 thing NV released a long time ago said on it 1tri/clk setup rate, so unless they changed it...?

On the TMUs, didn't the same diagram show only 4 TMU/cluster? Or am I seeing things wrong...?
 
There are 16 LS units per cluster. A bilinear sample takes 4 loads so there are an equivalent of 4 TMUs per cluster and 64 TMUs per Chip.

I think you really can't compare it. Seems like NVIDIA introduced a much more flexible way of doing texture sampling in Fermi. Especially for anisotropic filtering there could be some possible optimizations if you go down to individual point rather than bilinear samples.
 
Fermi Fun Fact: Launch products will have less TMU's than GT200.

edit: edited for clarity! (after this post I'm going to know much flak I can take for trusting someone)

Haha, oh dear. Its not the number of texture units, its how you use them?
 
There are 16 LS units per cluster. A bilinear sample takes 4 loads so there are an equivalent of 4 TMUs per cluster and 64 TMUs per Chip.

I think you really can't compare it. Seems like NVIDIA introduced a much more flexible way of doing texture sampling in Fermi. Especially for anisotropic filtering there could be some possible optimizations if you go down to individual point rather than bilinear samples.

Something like what ATI did with 5870, or going even further? Or something completely different even?
 
Back
Top