this is impossible due to the end of the world schedule
Nvidia GT300 core: Speculation
- Thread starter Shtal
- Start date
- Status
this is impossible due to the end of the world schedule
The Battleship? Who's the H.M.S. Hood then?
If GT300 has 512 alu's as rumoured, then it might be r7xx vs gt2xx all over again. I dont think they will be able to increase clocks by more than 10-15%. Since amd just delivered a 2.2x increase in throughput, nv needs a minimum of 3x increase over gt200 to stay competitive. Otherwise, they will be gunning for a gtx 385 and gtx 395 all over again.
Well let's take the worst case scenario and consider D12U to be a GT200b on steroids + X11 compliance. Do the math with only a 10% increase in frequencies and the result is above the 2.6x mark at many spots. That still doesn't mean much since those are still sterile on paper numbers. The cream in the pudding about NV's coming architecture is if they've changed anything in the architecture and how much it could have increased efficiency from A to Z.
If the rumors are true and its roughly below or above 500 square millimeters and even if it's nothing else than a GT200 revamp, it still doesn't sound bad. The bad part about it is that it's going to arrive late for the time being. The rest is fairly unknown.
Except that it's pretty clear already that die size difference will be smaller.
How?
GT200 & RV770 on the same process is about 200mm2 in difference, the same difference can be seen with this GT300 & RV870.
GT200 and RV770 wasn't on the same process.
GT200b is 490 mm^2 (if my memory serves me right). That's 230 mm^2 between GT200b and RV770.
Assuming that G300 will be ~500 mm^2 the difference between it and Cypress will be ~170 mm^2.
170 is less than 230 and two times less than 340.
It's not all about flops
You can have a gazilion ALUs but if can't feed them you can only go so far. Let's wait and see.
Then it must have something to do with amd's architecture. From the same shader, they are unable to extract as much utilization as nv.
I'm not sure I quite understand that. Compared to rv790 throughput of the HD5870 is "only" doubled, so a gt300 with 512 alus and slightly higher clock should be faster (since a gtx 285 is already faster than a HD4890, this would in fact widen the gap). Though rv870 might have some clocking headroom for faster editions...
I lost track of all the rumours though, what memory / bus width is gt300 supposed to be using? The HD5870 will supposedly be a bit limited by memory bandwidth, so if gt300 does better there that could also make it faster.
If nvidia stays with similar shader organization, I think reaching decent clocks is key to achieve competitive performance. A big part why gt200(b) failed to impress was its relatively low (shader) clocks compared to g92(b) IMHO, so despite the supposedly higher efficiency performance per area actually went down (ok the DP unit also contributes to this, but with clocks like the g92 the whole series would no longer really be competing with rv770/rv790).
I was going by the latest revisions; G200b (480mm2) & RV790 (290mm2) with a difference of just less than 200mm2.
if the Crysis bench on HD5870 is correct, no, it wouldn't be faster than HD5870. in fact it would be a fair bit slower.
OpenGL guy
Veteran
You could also reference this thread.
- Status