Monday, March 12th 2012
GK104 Silicon Roughly As Big As G92b
When the first PCB shot of the GK104 reference board surfaced, it sent the punters estimating the die area of the GK104 GPU, which the pinned at somewhere around 320 mm². A newer close-up picture of the GK104, helped calculate the figure more accurately, down to around 300 mm² (±5 mm²). This calculation also takes into account that the GK104 chip-package is as big as that of the G92, and the die just as big. It was compared alongside a 55 nm G92b chip. GK104 is NVIDIA's newest performance GPU built on the Kepler architecture, on which SKUs such as the GeForce GTX 680, are based.
Source:
Expreview
34 Comments on GK104 Silicon Roughly As Big As G92b
In an ideal world we'd see Nvidia continue their original plans and price their highest end GK104 around $300, offering performance on par or better than the HD7970, but I have my doubts.
Because of the highly tailored shaders that seems to work wonders on games set up to take advantage of them, ie games that are in TWIMTBP then nVidia are thinking they can overclock the 660TI and call it the 680, then cripple the others to make the slower cards.
If the rumors are to be believed then the GK104 will destroy some games (the optimised games') but in 'fairly' developed games AMD's 7970 is the more powerful card.
Once again it comes down to price but also in this round it what games you play and whether nVidia 'help' code them!
In any case it's a lot smaller than first reported and without knowing anything else my performance expectations are certainly lower now. An almost 20% smaller die, and assuming same perf/mm^2 or perf/transistor, is going to produce a chip that is 20% slower than I first expected. If GK104 is faster than Tahiti even if only by 5% or 10% it would be very remarkable IMO, considering Tahiti has been measured as 360-380 mm^2, and considering from where Nvidia came from.
If W1zz could travel half-way across the world to get home, build up a rig for the card, get it tested, and write a proper review in like 3 days....man, he'd have to not sleep since starting!
But that laso hleps put things into perspective...W1zz says nVidia gave him a chocolate cookie, and posted a pic. The NDA must expire relatively soon, but I cannot imagine that nV wouldn't give reviewers a week, at least, to get a review done. That puts things at some time next week, iMHO.
Of course, Z77 and IB is supposed to launch soon too, or something...and i haven't the foggiest idea when that is athough I'm supposed to be TPU's board reviewer. Clearly I'm not doing any launch reviews there, nor do I have access right now to unreleased hardware, so perhaps my judgement on NDAs and the time frame given to deal with publishing a reivew is not accurate. Oh well.
GTX570 for $250 local, on that day? That'd be awesome.
G92B stuck around for quite some time. I hope this GPU does as well.
The question I will want to have answered when this finally makes it "out from the wild"; did Nvidia have this figured into the chip early in the design or is this just a way of making a "silk purse from a sow’s ear"?
If Nvidia has really applied the idea from the get-go they now might be able to provide granularity, response and speed to implement smooth transitioning while in actual game play. Although, if an afterthought once they saw the GK110 was gimp, and the only other player was to call-up the "B-Leaguer" and pump it with steroids in the hope it can become proficient at every play and be the "game changer".
I really think that Nvidia is working from, "nobody normally detects over 60-70 FPS" so we'll implement a throttling device that maintains that FPS by vary the power/clocks to limit that. It has merit as like a racecar on a wind’e track excessive HP can be hard to apply, but a car with "enough HP" and then a nitrous bottle to provide the quick boost when it needs to stretch its legs on straight-aways is a threat. When it's not a endurance race then nitrous bottle has advantages, though long term just good well modulated HP is the stamina to do it hours on end and not wreak havoc mechanically.
If this is just a face saving measure by Nvidia and it works and a true a "game changer", will Nvidia provide the added reward and price it compensatorily to really put AMD in a bind, we wait...
Nvidia have done for last gens a big fat chip with a fat memory bus thats doing the same thing as the far cheaper design of amd.
HD4870X2 was still smaller(both chips combined!) than a GTX280 if i remember correctly.
nvidia must have done A.\ something big. B.\ overclocked it alot like amd did not do.