Nvidia boggles the mind with the GTX670

Okay, so its review and launch day for the GTX670. If you aren’t yet familiar with Nvidia’s new generation of graphics cards, read through my GTX680 Analysis to understand what’s different in the new chips and how they operate. Three cards have been launched so far; the GTX680, the GTX690 (analysis still coming, some interesting things pop up) and today’s offering, the GTX670.

Like the GTX680 and GTX690, the GTX670 is based off the GK104 design, which is essentially an upgraded GTX460 with parts chopped off, parts added together and parts made smaller to make today’s Frankenstein. The chips boast a smaller node process to 28 nanometers and are much more efficient than their predecessors in the outgoing Geforce 500 series. But what’s really interesting that we’ll see in the coming months is exactly how Nvidia plans to differentiate their chip lineup. After all, today you’ll see Nvidia’s equivalent of the HD7950 – a card performing a hair’s breadth away from the flagship, costs far less and once overclocked takes the performance crown easily.

But first, the interesting thing to take away here is that TSMC is still having yield issues with Kepler and the 28nm process. Nvidia’s actually gone and re-branded their GT540 parts as the GT640, producing a few different versions but having two of them – the GT645 and the 144-core GT640 – manufactured using the 28nm process to allow Nvidia to sell off dies that would have otherwise been tossed out. The breakdown between TSMC and Nvidia gets more tense because there aren’t any known reports of other manufacturers having fault with 28nm yields. Either their designs are more fault-tolerant, or something is wonky with getting Kepler tapered out properly.

One shader disabled, the rest left intact.

Whatever the case may be, today is another paper launch. Nvidia still can’t get enough GTX680s out the door and into the right regions. Yes, you read that right, regions. There are tons of shops online and in other countries that have stock of the flagship, but aren’t selling either because of high prices, AMD’s bundle of three games with their HD7000 series or simply because the customers are waiting for the GTX670. Well you’ll have to wait a bit longer, as TSMC’s assumed outputs of Kepler units per month is only 10,000. What you will see today isn’t a different card that would have been a GTX670, its actually a GTX680 with a disabled shader module.

Looking at the chip layout, the GTX670 ships with 1344 CUDA cores and 112 Texture units, making up seven out of eight possible shaders, the extra one being fused off by laser. I expect that these weren’t perfectly functioning GK104 units and Nvidia changed their plans for the lineup in the interim. If you’re getting one of these puppies, it’s best you save up for one now before Kepler’s yield issues are sorted, as Nvidia may change the chip later for whatever reason. The GTX670 is now only a slightly slower GTX680 and fits in the same space in the market, fighting off the onslaught by the HD7950 and HD7970 for the high-end and enthusiast segments. Clocks stand at 915Mhz for the core and 6Ghz for the memory.

But anyway, onto the card itself. Its covered by a dual-slot cooler and features a blower fan on the far right exhausting all heat out the vents. But wait, that PCI-Express power slot in the middle HAS to be a shop, right? I mean, that’s completely the wrong place to put it. Yeah, if you flip the card over, the PCB is dwarfed by the 10″ shroud, measuring only 6.75″. Disabling the shader module allowed Nvidia to adjust core circuitry and focus on cooling the card and getting power consumption down (remember what I said about this being a faulty/disabled GK104 chip? Looks like I’m right). What’s really impressive is that with a shortened PCB to begin with, third-party OEMs might figure out a way to produce a single-slot and much longer GTX670.





The card requires two six-pin PCI-Express power connectors and has a typical power draw of 141watts. This card will end up being the overclockers choice because the board now has a maximum power limit of 225 watts including the 75w that the motherboard provides through the PCI-Express slot. With extreme cooling and more aggresive BIOS and clock settings, I can see quite a few vendors creating their own version of the GTX670 in a “Ti” guise. The back of the card reveals two dual-link display ports, one full-size display port (pity, should have been a mini-connector) and a HDMI 1.4a port. All four can be used to drive a screen each for Nvidia’s Surround feature. This previously required two cards in SLI to achieve a three-monitor setup.

Moving onto the benchmarks, its apparent that the disabled shader doesn’t reduce performance that much, even though it also cuts power consumption down by 30w. In Battlefield 3 the GTX670 finished within a ten-frame margin behind the GTX680 and you can see it easily showing the HD7970 its tailpipes. Running at native 30″ resolution, it sticks even closer to its bigger brother to within 5fps, drawing equal with the HD7970. Its the same story with Crysis 2, drawing up with the competition in DX9 and blowing past them in DX11. Its a really tremendous thing to see, and Nvidia must be commended on their design of Kepler – even with the loss of a limb, it still keeps up. The card loses out at 30″ resolution in Crysis 2, but this is more to the HD7970 having a bigger framebuffer.






I’m only going to show 30″ resolution tests for now because all other games produce 70+ frames per second at 1080p. What you saw in Crysis 2 and Battlefield 3 basically plays out everywhere else. Performance in Metro 2033 splits the cards by less than four frames – was that extra shader core being unused in games? In Skyrim the card draws up again next to the HD7970 and the GTX680, showing playable performance even with FXAA enabled. DiRT3 gets close to the magical 60fps barrier, easily beating out the HD7970. DiRT2 favoured AMD cards, but this time round the tables have turned. WOW Cataclysm finally gives the HD7970 the boot, clearly showing that the game isn’t optimised for GCN yet.






So now, you may be wondering what kind of knock Direct Compute performance has on the GTX670. I mean, if shaders aren’t used in the game that’s fine, but if CUDA cores aren’t being fully exploited, then that’s another thing altogether. Look below the the MediaEspresso benchmarks where video files are recoded into a high-quality clip that an iPad could play. Recoding to MPEG2 shows Kepler-based cards finishing the recode in half the time of AMD’s offerings. Changing to H.264, the GTX670 draws up neatly with its bigger brother, producing idential results. Anandtech’s Civilisation V Compute benchmark shows the GTX670 only slightly behind the GTX680. Honestly, for most games and applications, it looks like disabling that core didn’t much damage the card’s ability to perform well.






Finally, moving into temperature and power consumption, the card keeps temperatures in really well, idling at 41 degrees and maxing out at 80 degrees on load. Power consumption is reigned in , roughly matching the HD7950 for performance. You’d only need a decent, stable 550w power supply at this point to power the beast, just like the GTX680. What’s really going to boggle your mind, though, is the SLI scaling. That, right there, is GTX690 performance. Two cards that are cheaper than two GTX680s, hanging with Nvidia’s fastest card to date. Even dual HD7970s stand a chance, the scaling is just phenomenal at an average of 80% improvement across the board.





So, here’s the drill. Its cheaper and better to get the GTX670, overclock it and have GTX680 performance for approximately R1000 less.  Even if it sells for R4500 here, you’d be mad to choose the HD7970 over this because it wipes the floor with it, all things considered. Those of you considering a dual-GPU solution, you’ll only need a 700watt power supply to get GTX690 performance for less than you’d pay for an actual GTX690. This card will also effectively cannibalise GTX680 sales unless Nvidia reveals that drivers weren’t optimised and bla bla bla and fixes everything. The card is crippled, for heaven’s sake, and it still plays tag with the flagship costing considerably more. And uses more power.

But, there’s a crucial flaw in Nvidia’s plan for dominance and that’s TSMC. Without a supply, a healthy one at that, of 28nm dies and good yields, their entire lineup of Kepler’s derivatives will be starved. Obviously they’re going to focus the majority of development and supply to the GTX660 (when it arrives and kicks the HD7870 aside) and the GTX670. Only ten GTX690’s arrived in the country this week. Ten.  Those are using good dies that could have been put to better use in the GTX680 or even the 670 to give fans more than another stupid paper launch.

I said before the reveal of the GTX690 that I wondered if Nvidia had developed some common sense what with all the problems they’re having. Giving gamers in the mid-range a good supply while they worked on the real flagship would have been just fine – AMD’s HD7990 is only showing off at Computex, after all. When the GTX670 lands, it’ll likely sell for much higher than its suggested retail price and stocks will be limited.

A paper launch for what will become the gamer and overclocker’s card of choice? That deserves  an honorary mention, not an award. The recommended retail price is $400 (approx R3244), so all that’s left is to wait for the obligatory price drop from AMD in the coming weeks. And we’ll probably see the GTX660 with another shader module disabled in three or less weeks.

Source: Tom’s Hardware, TechpowerUp!, Anandtech

Discuss this in the forums: Linky