Table of Contents
It’s been a few months since the AMD camp released its complete lineup from enthusiast down to budget cards with no reply from NVIDIA on the lower end of the spectrum. Starting today, that will change. NVIDIA has finally brought to the table answers to the 7870/7850 and 7770/7750 cards in the GTX 660 and GTX 650 family. As usual, board partners put their own spin on the reference cards in the way of souped up coolers, hardware, and factory overclocks.
Galaxy has done the same with their GTX 660 GC 2GB GPU offering as well. Below we will take a look at the architecture of the reference cards and dig a lot deeper into Galaxy’s improvements on them. Let’s see if NVIDIA has what it takes to compete with AMD in the mid-range/budget arena and what Galaxy has done to improve on the reference design.
The first item we will take a look at are some generic features for the GK106 GPU used in the Galaxy GTX 660 and presumably all 660’s then take a high level look at its architecture. Some things to note here are the 960 CUDA cores versus the 1344 for the 660Ti and the 80 Texture units versus the 112. For the GK106 core, the 960 cores are the full allotment, this is not a cut down version. Things only get watered down from here for the GTX 650.
On the memory side of the house, the 660’s 2 GB of DDR5 Memory runs at 1502 MHz (6002 MHz DDR5) resting on a 192 bit wide bus (144.2 GB/s bandwidth). For those that did the math, and are left scratching your head, there is more on that oddity below the specifications table.
The GTX 660 has a thermal design power of 140 W with a thermal limit of 98° C (toasty!). Reference clocked models run “typically at an average of 115 W in non TDP applications” and, when overclocking and maxing the power slider out (a mere 10%) to 127 W.
Taking a look at the innards, the GK106 core is divided up in to three Graphics Processing Clusters with a total of 5 SMXs making the 960 CUDA cores as mentioned earlier. You can also take away from this picture that the memory is divided up into three controllers making up the 192 bit bus (normally 1.5GB of VRAM). In this implementation, and select past NVIDIA GPU’s, they run mixed density memory modules to reach the 2 GB quantity. In this card, you have a 128 M x 16 (1 GB, on four memory IC’s), and two 64 M x32 configuration on the remaining four IC’s (512 MB each) to reach 2 GB. NVIDIA has chosen to give you more ram in favor of having matched bus speeds across the controllers. I’m not sure if there are any limitations on this setup, though. I suppose that could show at high resolutions using a lot of the VRAM though.
Below are some benchmarks completed by NVIDIA. Settings varied by game of course but the resolution was 1920×1080 for all of them. These are marketing slides so read into them as deeply as you wish.
The other release today will be the GTX650 which is of course a bit lower on the totem pole. The 650 takes a drop in CUDA cores coming in at 384 (presumably with 2 SMX’s). With that however the core clock defaults at 1058 MHz. This card has 1GB of DDR5 memory running on a 128bit bus at 1250 MHz. Below are the NVIDIA specifications for that card and their performance benchmarks NVIDIA ran. It looks to beat out AMD’s low end card in the benchmarks used. I wonder what the price will be for it? Seeing as how they put it up against a 7750, it should be around that pricing, perhaps a bit more.
NVIDIA has targeted these cards as “Kepler for every gamer”. Through various outlets like Xfire and Steam, they have crunched the numbers and found a graphics card sweet spot for price point of around $199-$249 (anyone remember the 8800/9800GT upon release?). I personally remember the release of the 8800GT and the price for performance chord it struck in the hearts of many enthusiasts, including myself. The GTX 660/650 is aimed to please those same users again.
Okay, enough about the generic stuff, let’s see what Galaxy managed to put together and release in their GTX 660!
Packaging & Accessories
This is our first review with a Galaxy supplied card and is the first time in years I have seen one in person. One of my first reviews for Overclockers.com however was the venerable Galaxy GTX 470 so I am no stranger to the brand. That review being two years old and a generation removed from Kepler, things of course do change. Gone are the box like patchwork of metal panels on the coolers for something a little more sleek among the aesthetic changes.
On the front of the packaging you can see Galaxy has their “G” symbol prominently on the front showing the name of the card, the 2 GB of vRAM, the fact that it is overclocked (the “GC” nomenclature) as well as the custom cooling. On the flip side of the box are more marketing and features, same with the top side. In the last picture we finally get a peak of what is inside. The GPU rests inside the upper portion of the box in a anti static bag securely. Below the card is where the accessories are stored.
The accessory stack included with the card is fairly minimal, but has what is needed to run the card in a driver disk, molex to 6-pin PCIe adapter, DVI to VGA adapter, along with a manual and quick start guide.
And here she is, the Galaxy GTX 660 GC 2 GB. One of the first things you notice are the two large fans and three pipe cooling this model offers over the reference cards. Galaxy has put some thought in to the fan blades as well shaping them to ‘reduce noise output through its entire operating range’. The design of the shell is said to help force air out the rear of the case. I’m sure it helps, but with it not being sealed as is the case in most reference, blower-type coolers, some of the air escapes in the case. Any effort minimize heat dump inside a case is certainly welcomed though.
Moving on to the back of the GPU, typically there is nothing to mention, however Galaxy has made an attempt to cool the MOSFETS a bit better with holes in the PCB to allow for better cooling in that area. I’m not entirely sure how effective that is, but it wouldn’t appear to hurt anything at minimum. They have also chosen to place some memory IC’s on the back of the card as well.
One thing I was surprised to see on a mid-range level card are voltage read points. They are not in one specific area you can see the small, labeled orange circles scattered about the rear. The memory and ground are towards the top/middle to the right of the memory IC’s, while the GPU is towards the left side of the board. It also measures PCIe voltage and the 5v feed its getting. The last thing of note here is the single SLI tab allowing one to SLI two of these cards if they choose for increased performance.
Moving to the output area of the card, you can see Galaxy is using a typical setup of HDMI, DisplayPort and two DVI-d ports; plenty of options to hook up to your monitor(s). If you still using the VGA input on some monitors, they do include a DVI to VGA adapter as mentioned above. One nice feature to prevent interference is that the DVI outputs are shielded.
Stripping the card of its heatsink and fan, the factory thermal application was sound. In the first picture you see the three heatpipes and fin array used to cool the card. Also exposed is the 6 phase power (versus 5 on the reference version), which helps deliver cleaner power to the GPU and memory. Outside of that, you can see dual holes for mounting coolers to it which should expand the amount of aftermarket coolers and water blocks available to be used on the card if that tickles your fancy.
Last I will show a closeup of the new GK106 core that is under the hood as well as the Hynix H5GQ2H24AFR R0C which comes in two flavors; at 1.35v its rated for 4 Gbps while at 1.6v 6 Gbps. Checking on the voltage its set to 1.6v giving a rating of 6 Gbps or 1502 MHz (its stock speed).
Overall I like the look of the card. Galaxy’s past shrouds were a love/hate thing, and they have improved this card’s sheath.
Performance and Overclocking
- Intel i7 3770K CPU @ 4 GHz (Overclockers.com Approved!)
- MSI Z77 MPower (Overclockers.com Approved!)
- G.Skill RipjawsX 2 x 4 GB 2133 MHz CL7 1.65 V @ 1866MHz 9-9-9-24
- 60GB OCZ Vertex 2 SSD (Overclockers.com Approved!)
- Seasonic 1000 W
- Galaxy GTX660 GC 2GB (1006/1097 boost stock – 1080/1170 boost overclocked )
- Windows 7 64 bit Operating System
- NVIDIA 306.23
- All Synthetic benchmarks were at their default settings
- Unigine Heaven (HWbot) was run with the “extreme” setting
- Alien vs. Predator – 1920×1080 with highest settings offered (4x AA, textures set to highest)
- Batman: Arkham City – 1920×1080, 8xMSAA, MVSS and HBAO, Tessellation HIGH, Detail Level: Extreme
- Battlefield 3 – 1920×1080 at Ultra settings (4xAA/HBAO by default)
- Dirt 3 – 1920×1080 with 8x MSAA and all settings enabled and at Ultra where possible
- Metro 2033 – 1920×1080, DX11, Very High, 4x MSAA/ 16x AF, PhysX OFF, DOF enabled, Scene: Frontline
- Civilization V – 1920×1080, 8x MSAA, Vysnc OFF, High Detail Strategic View: Enabled, Other Settings: High, using full render frames value ( / 60)
- More detail is in our article: Overclockers’ Updated Video Card Testing Procedure
Below is a screenshot of GPUz 6.4 and what it shows about the card. As you can see it agrees with the basic specifications listed from Galaxy, so all is well there. Note the boost listed is the average boost. With the power limit at default, the boost topped out at 1090 MHz, nearly a 100 MHz jump.
As far as GPU overclocking software goes, Galaxy has added an application to the fray. Named Xtreme Tuner Plus, like all other software of the like, you have a nice display showing some basic stats and current settings. You are also able to change your power target, core clock, memory clock, GPU voltage, memory voltage on certain cards (not this one so far), as well as your fan speed. What is nice about this on the same screen, you can adjust your fan profile. I like this as it saves you a click or so compared to other software where its inside a tab occasionally requiring an unlock to access it. Another nice feature that is common throughout most applications like this is its ability to display/monitor things on a graph (GPU temp, speed, memory speed, fan speed, etc) when you click on the monitoring button. Another window pops out and displays the graph.
Overall the software is easy to use and accomplishes what it needs to do. One item to improve on this software would be to make the application take up less real estate on your screen. Its not huge or unruly, but I think it could be a bit smaller without losing readability.
Ok folks, let’s look at some benchmarks and see whats going on here. The first we will look at is 3DMark 03 and 3DMark Vantage. Although 3DMark 03 is a little long in the tooth, it still shows significant gains between all cards. What is a bit surprising here is the very low score out of the gate in 03, posting a 81.6k out of the box. With that score, the GTX660 is losing to its cheaper competition, the 7850, and a 7870 by a significant amount while trailing its bigger brother, the 660ti by not nearly as much. Since darn near every Kepler based card is losing out to AMD’s implementations on this bench, I think its fair to say Kepler just doesnt play nice with 3dMark 03.
When we look at the 3DMark Vantage results, the story moves back to what was expected. The Galaxy GTX660 put up a solid showing here with a 26.6k result at stock speeds. This bests the 7850 as expected, and also trumps the 7870 as well. It falls just 10% behind the 660ti. Not too bad at all.
The next set of benchmarks we will look at are 3dMark 11 and Unigine Heaven (Hwbot, extreme setting). Both of these benchmarks make heavy use of tessellation and are a pretty good barometer of modern games features and performance. As you can see below, the GTX660 put up 7,004 point in 3dMark 11 and manged to beat out the 7850 by 13% and the 7870 by almost 9%. Again, an impressive showing.
Moving over to Heaven, we see more of the same really with the Galaxy 660 beating out the 7850 by about 11% and the 7870 by almost 6%. NVIDIA’s 660ti held a comfortable lead throughout this testing.
The part I was most curious about is seeing if the lead the Galaxy GTX660 held in the benchmarks translated to gaming. First up is Alien vs. Predator (DX11). In this bench the 660 managed 43.7 FPS beating out the 7850 by over 6% or almost 3 FPS. It did fall short compared to the 7870 by a little over 1.5 FPS or 3.5%. Hardly noticeable, but an expected result none the less.
Moving on to Batman: Arkham City, we see an anomaly, but in a positive way here… the GTX660 matches or beats out all cards used for testing. It easily beats the 7850 and 7870 11.5% and ~5% respectively while it matches FPS with the 660ti. I confirmed the benchmark settings and reran the test to confirm my results and they were accurate. After looking around the only thing that is different are the drivers. In the 306.02 release and on, Batman: Arkham City receives an “18%” increase with those drivers. Looks like that is what we are seeing here.
Last in this graph is everyone’s favorite First Person Shooter, Battlefield 3. Results here fall in line with expectations really with the Galaxy card posting up 55.4 FPS beating out both the 7850 and 7870 in this title by a few FPS. Outside of the benchmark we use, in multiplayer, I was able to run this game on Ultra settings, using FXAA and SSAO with no problems what-so-ever at 1920×1080. I tried with the Ultra defaults, but I would bottom out a bit lower than my tastes so I needed to make that quick change I noted for seamless game play. That said, making that switch you would be hard pressed to notice any IQ differences between the two. A different website completed some testing on just that.
The next graph shows Civilization V, Dirt 3, and Metro 2033. in Civ V, the 660 manages 57.3 FPS matching the 7850 and falling short of the 660ti as expected. In Dirt 3, we see a solid performance here out of the Galaxy card posting 72.7 FPS beating out the 7850 easily and within 1 FPS of the 7870. The last game we test is Metro 2033, the GPU killer. Here the card performs within one FPS of all cards listed coming it at 24.4FPS.
Cooling and Power Consumption
The dual fan, three pipe heatsink solution Galaxy used performs quite admirably. Below 60% fan speed (60 °C with the default fan profile) their system is pretty quiet and cools quite well even when overclocked. It’s not until you ramp up to to the 80-100% that the fans start to make their existence known. Even then it’s not too bad, but I don’t know of many people outside of benchmarking that would crank the fan like that, again with this card its just not needed. Below are the results using the default fan profile.
Taking a look at the power consumption, we can see the expected outcome. As I have noted in past reviews, its been a long time coming to see this type of performance with such little power consumption. This card peaked at 256W at the wall while averaging around 240W or so (overclocked) in the 3D Mark 11 combined test.
Pushing the Limits
Due to the +10% Power limit imposed on these cards, I was not able to push much more past the 24/7 overclocks I listed. In fact, I wasn’t able to touch the core hardly at all because it would throttle back the core clocks. That said, the boost clocks for these benchmarks hit around 1170 MHz without throttling. I pushed the memory a bit and came up with these results with a 4.8 GHz CPU pushing the card. Not too shabby, but please lets get this power limit raised on this card ASAP so we can clock it more without throttling!
The market has been begging for NVIDIA to come out with their low/mid-range cards for a while now, and finally they have. The reference clocks/specifications show this card is a little powerhouse and the benchmarks show that as well. You are not going to be able to play modern games at ultra settings with copious amounts of AA, but in a lot of titles you will get close. For example, BF3 at 1920×1080, ultra settings with FXAA and SSAO instead of MSAA/HBAO played fine. In some heavy action I bogged down just a bit, but not even enough to annoy me. Better yet, play it on HIGH settings with SSAO and the card doesn’t really flinch. What Galaxy has done to this card – a better PCB/power delivery vs reference, an aftermarket cooler, and an overclock out of the box – only serves to improve what I believe to be a solid base.
The cooling solution keeps the card nice and cool. The temperatures with the default profile do not reach the 70 °C boost throttling point so you can set it and forget it for the most part. Notice I said for the most part, the concern I have with this card will only effect those trying to push hard on overclocks really, but its the power limit. It’s set to 10% maximum. With it being that low, I was only able to manage a +72 core before it hit the power limits and scaled back. While I didn’t notice this in games, it is apparent in benchmarks. This isn’t a shortcoming in the software or the hardware (remember Galaxy has beefed up the PCB, but in NVIDIA blessing this higher limit. Galaxy is working with NVIDIA on raising this limit as I do believe this card has more left in the tank.
Pricing for this card is said to be at an MSRP of $239.99-$249.99. The price point of this card places it right between a 7850 ($199-$244) and the 7870 ($259-$379). It’s performance in the games/benchmarks tested here, eclipsing the 7850 with relative ease in all but 3DMark 03, and on par with the 7870. Galaxy has a nice package with their GTX660, giving users a superior cooler versus the reference model, superior PCB for stability and potentially better overclocking, and functional overclocking/monitoring software Xtreme Tuner Plus. Galaxy has a winner on its hands with their GTX660 GC 2GB card.
~Joe Shields (Earthdog)