About 2 weeks ago Nvidia updated it’s Facebook page with an image of a the unnamed graphics card and the teaser “IT’S COMING”:
A week later we found out what “it” was. A GTX 690 – essentially 2 GTX 680 GPUs and their RAM on a single PCB. I have to admit, I was somewhat surprised by this. At the time the GTX 680s were hard enough to find due to GPU shortages (supposedly due to manufacturing difficulties with TSMC,) so to try to put two in one graphics cards seemed a crazy move. On top of this, The real money in graphics cards is made through the bulk sales of cheaper “mainstream” cards rather than the smaller sales of top end cards. The real kicker in the announcement though was that the GTX 660Ti wouldn’t be arriving until H2. I assumed this meant all the lower cards would be late too…
Just one week later, came the graphics card I was waiting for – GTX 670. The GTX 680 is a great card, but is too expensive to justify. The GTX 670, whilst still a high-end card, is some £100GBP/$150AUD cheaper. On to the hardware.
GTX 690 is easy to summarise. As I said before, it’s essentially 2 GTX 680s on one PCB. The GPUs on this board are clocked a little lower but could still boost to clock speeds higher than the 680’s base clock under the right conditions. The card reaches around 95% of the performance of a pair of GTX 680s in SLI, whilst consuming far less power, generating less noise and less heat. Heat. Herein lies the weakness of most dual GPU cards – a centrally mounted fan exhausts the heat from one GPU out the rear of the case, but exhausts the heat from the second into the PC case, where it will end up raising temperatures of other components. The card costs a similar amount to a pair of GTX 680s so this would be a great route to go if you were planning on diving straight in with two GTX 680s.
GTX 670 is based on a cut down GTX 680 GPU. In my post on the GTX 680 I questioned how Nvidia was planning on scaling it’s GK104, and now we know:
Rather than cutting out a GPC, Nvidia has disabled a SMX within one of the GPCs. This means a reduction in stream processors of 12.5% (1344 rather than 1536). On top of this the GPU runs 9% slower (915MHz rather than 1006MHz). Rather surprisingly, this graphics card only runs 5-10% slower than the GTX 680, this is due to the aggressive GPU boost settings which can boost the clock speed by 18% rather than the GTX 680’s 10%.
In the GTX 670, we have a card that is about 25% cheaper than the GTX 680, but only 5-10% slower. Sounds like a bargain to me and will likely make it’s way into my next PC.
So, it’s been a week now since Nvidia released the GeForce GTX680 and and response has been overwhelmingly positive. AMD have had a good run by themselves at the top end in the 3 months since they introduced their 28nm part in the form of the Radeon HD7900 series. These boards were impressive and many (myself included) wondered how Nvidia were going to compete after their last 2 top end cards were pretty hot and power hungry. The answer is to get rid of the shader clock and pack in a LOT more processor!
Previously, Nvidia ran its shader clock at twice the speed of the rest of the GPU but Kepler has done away with this approach. In doing so they have decreased the power draw and temperature the GPU operates at. Obviously with this halving of speed there will be a massive decrease in performance. In order to counter this, Nvidia has tripled (yes, tripled) the number of stream processors to 1536, as well as increasing the clock speed by nearly 30%. Unfortunately, the RAM side of things has seen no such overhaul. Whereas AMD’s 7970 has a memory bandwidth of 264 GB/s, the GTX680 has just 192.3 GB/s. This is actually down on the GTX580, albeit by only 0.01GB/s. With such a powerful GPU this could be a problem, if only minor. The situations when you need the most graphics grunt are at higher resolutions (surround for example), in 3D, and when applying heavy AA. Guess what? Every one of those scenarios also requires a decent amount of graphics card RAM with plenty of bandwidth.
Most websites have found the GTX680 to be noticably faster than the 7970 in the majority of games (obviously it depends on the game engine, developer etc.). The real issue (for AMD) is that Nvidia have launched this card cheaper than theirs. We can expect price drops on at least the top tier of AMD’s range, and possibly further into the upper-midrange. This lower price will have been brought about in no small part by the fact that Nvidia have produced a chip that is physically smaller than AMD’s. for the last couple of generations Nvidia have produced GPUs that have been, frankly, huge. The GTX 580 was a whopping 520 mm2 this compared at the time to AMD’s 6970 at 389 mm². Big difference. This time around however, whilst AMD is producing a 7970 chip at 365 mm2, Nvidia have managed to squeeze all thGTX680’s power into just 294 mm2, allowing the use of less silicon, less cost to manufacture, and less price to the end user.
The next big question is the mid-range. How will the new GF104 scale, and how is Nvidia planning to do so. We should be finding out in the next few weeks, but if the GTX680 is anything to go by, we could have some cheap powerful chips hitting the midrange.