Ya, what about it. I didn't miss the point at all. NV can easily release a card for $399 with 20% more performance than a GTX580. That would make HD7970 at $549 insanely overpriced
Some might argue that the GTX 580 and 7970 are already insanely overpriced - you don't need an unreleased card to see that- a $260 unlockable HD 6950 makes that abundantly clear.
At stock speeds the HD7970 is barely faster than an HD7970.
Hardly surprising
You keep focusing on 2GB of VRAM limitation when for 99.9% of people that's plenty fast
99.9% of people don't use enthusiast level graphics cards...and what I wrote was "
Why would Nvidia go from 384-bit to 256-bit for their top GPU ?" -the statement was in reply to your fixation with GK110. You really think a GK110, like any other enthusiast card, will be purchased in any significant numbers. More to the point, are you expecting the GK110 to have a 256-bit memory bus?
(What is it and Guests with straw man arguments?)
It's also why I said "
I really don't see it happening with those specifications"
IF the card being talked about is the GK104, then the specifications being bandied around are a 40MHz lower core clock than the 560Ti, no shader hot clock, the same bus and framebuffer, smaller die, better performance/watt, better performance/mm² for a nominal
55w increase in TDP
but lower temps and a 80% increase in performance.
Usually when something seems to be too good to be true, it's because it is. The flip side of this is that numbers like this wont be in any way a good thing for consumers if true. If the GK104 meets or exceeds 7970 performance, rest assured that Nvidia will price accordingly.
HD7970 can't take advantage of > 2GB of RAM since it's not fast enough in those situations in the first place. The few games that use a lot of VRAM like Shogun 2 destroy HD7970 at 2560x1600 with AA.
Whatever...Just out of interest,
check the relative performance of the same 3GB 7970 with the 1.5GB version of the GTX 580 (albeit with a more powerful rig)
You are also assuming that 256-bit memory interface is a problem. You aren't considering that NV can simply increase TMUs, SPs and ROPs and match the bandwidth of the GTX580 with faster GDDR5 chips on 256-bit interface). They can squeeze 20% more performance from GTX580 without increasing bandwidth.
Yep. Probably why I wrote "
unless Nvidia have made a fundamental leap in GPU design that mitigates the reduced bandwidth"
They may have also moved to a much higher bandwidth GDDR5, they might also have improved their memory controllers, they might also have managed to pare away latency, they may have simplified the GPU by removing the double-precision element., and they may -as you've said- move to increasing ROP's and TMU's (maybe 48 and 96 ?)...which probably constitutes "
a fundamental leap in GPU design"
Bandwidth of GTX 580: 384-bit / 8 x 4008 MHz effective = 192384 MB/sec (usually expressed as 192.38GB/sec
Bandwidth of GK104...: 256-bit / 8 x 6012 MHz effective to reach 192384 MB/sec...a 50% increase in memory speed is required, and some 500MHz faster than AMD's GDDR5...the guys that pretty much invented the GDDR5 spec..
So basically the memory bandwidth limitation and "only" having 2GB of VRAM are only problems in your mind. AMD released a card for $550 that's only 20-22% faster without seeing what the competition can bring. I have no doubt that Kepler GK110 will be at least 40-50% faster than GTX580, which means GK104 should have no problems at all matching HD7970 at a much lower price if NV chooses to be aggressive with its pricing strategy.
A couple of points:
1. GK110 isn't here. This thread isn't concerned with what will in all probablility be a 384-bit + (possibly 512-bit) GPU. If 256-bit/2GB "isn't a problem" in your mind, do you think it likely that GK110 will be 256-bit/2GB ?
2. Nvidia has seldom been aggressive with its pricing strategy- unless responding to AMD's pricing. Haven't you noticed that AMD and Nvidia have been dovetailing price and performance since at least 2008?
3.
This is what VRAM limitation looks like at 5760x1080...and remember that at this price point there are going to be a few people might want to use this res or higher. Bear in mind that Charlie mentions that the GK104 includes DisplayPort, so if Nvidia are moving toward supporting single card 5040+ resolutions (and I think they must look to match AMD sooner or later) then a larger framebuffer (esp with/AA enabled) is probably a must- at least for the top tier card.
4. The HD 7970 has already demonstrated an ability for a significant percentage of cards to clock in excess of 1200MHz core and 7000MHz effective memory on stock cooling. You don't think that AMD might take advantage of this fact and bin for a 7980 (or whatever) if they need to sometime between now and when the GK110 launches ?
That kind of puts paid to the vast theoretical GK110 advantage...and that doesn't take into account that AMD might 1. revise/refine the GPU design, and 2. Have the HD 8000 series out by the time the GK110 drops....or are you privy to the launch dates as well as performance figures?
EDIT:
FWIW here's Dave Kanter's take just posted at B3D