Nvidia BigK GK110 Kepler Speculation Thread
- Thread starter A1xLLcqAgt0qc2RyMz0y
- Start date
-
- Tags
- nvidia
OTOH 320bit is definitely possible, nvidia doing it all the time (even if only as partly disabled parts). On a technical level a 320bit/40ROP/10SMX part would look quite sane to me, but I'm not sure it really makes sense, those markets might be served better with gk110.
I haven't seen any bandwidth scaling figures for these parts so I'd be cautious saying it must be bandwidth holding them back, though I'd agree adding more smx won't do much (just for reference, amd parts don't scale much with simd units for ages and it it wasn't only because of lack of bandwidth).
DuckThor Evil
Legend
Just snagged the Asus DC2T from newegg.
To play Diablo
Haha
Let us know, whether it boosts higher than 1201Mhz with stock settings.
It would definitely be interesting to see some bandwidth scaling figures especially for the 680. In any case on paper a 680 compared to a 670 has about 20% more ALU processing power and texel fillrates. However rasterizing and pixel fillrate is by roughly 9% lower on the 670, give or take the average performance difference between the two in 1080p/4xAA.
GK110 is on the other hand rumored to be roughly by up to 50% faster than a 680 in 3D (for which I assume the top dog is meant and not some salvage part). Assuming 1.5GHz GDDR5@384bits it's 50% more bandwidth, with a hypothetical core frequency of say 850MHz ~28% more pixel fillrate and assuming 5 GPCs a 5% increase in maximum theoretical rasterizing rate (or 27% with 6 GPCs); +58% texel fillrate and FLOPs.
Now my original point was that increasing theoretical rates on GK104 refresh by say 20% or more (through whichever way) won't obviously result into a =/20% performance increase if the increase in memory bandwidth isn't at least on a comparable level.
But dont they plan to increase the memory bus and ammount ? At least to a 384bits 3gb ? This seems at least a not so costly move from the GK104 and will position it ( marketing wise ) defacto in the higher position.... I dont see Nvidia going for a 384bits bus and 1.5gb memory after the 2gb GK104.
Last edited by a moderator:
He said 1.5Ghz not 1.5GB
lol my bad.
I'm wondering if more than 6gbps gddr5 is an option? Those memory companies "introduced" 7gbps gdd5 in 2008, maybe they can actually make it now. At least Samsung has a "-28" option for their gddr5, and while I can't figure out if that really stands for 0.28ns (which would be good enough for 7gbps it is presumably faster than the "-03" which I also can't figure out (only 04, 05, 5c are detailed on the shitty site) which is used for the 6gbps chips. Granted even at 7gbps that is not quite a 20% increase over 6gbps but it should help.
How do you know that?
We do? :neutral:
I think he's referring to OC'd models, but then again, Tahitis have their memories OC'd to around 1600-1700MHz too?
That of course tells nothing on how long or "well" they handle it, or how big percentage of the chips can handle it
I dunno sounds reasonable to me if clocks are reasonably high (so 50% more memory bandwidth, ~35% higher ROP-fillrate, ~70% higher alu/tex rate). That's assuming GPCs aren't a bottleneck (which clock-adjusted barely increase) nor that there are any other less obvious bottlenecks. Some of the compute stuff (not DP/ECC), while not area-efficient for gaming might also marginally help. The chip is quite a beast.
I've seen quite a few overclocks in that range.
So should we say Tahiti MC also does 1.6-1.7GHz, since many hit those speeds on OC, IIRC?
I must have missed where we were talking about Tahiti....topic was GK110 vs potential GK104 refresh fyi
Overclocking is indicative of results from typical material, not necessarily what a product is qualified or even designed to do.
Similar threads
