For the past year, Nvidia has engaged in an unusually extended launch schedule. It debuted its Maxwell architecture with a budget GPU in February of 2014, followed it up with the high-end GTX 980 and 970 late last year, and today, with its new midrange GPU — the GTX 960. While not as sexy as the high-end GTX 970 or 980, which retail for $330 and $550 respectively, it slips in at just $200, which makes it far more affordable and practical for a large number of PC gamers.
It’s also debuting into the teeth of AMD’s strongest current price/performance ratio. AMD has held a strong position in the $100-$300 price brackets, thanks to aggressive price cuts and the incremental improvements offered by GPUs like the Radeon R9 285. Can Nvidia break that lock? Let’s take a look.
http://www.extremetech.com/wp-content/uploads/2015/01/GTX960-Block-640x349.jpg
The GTX 960 is based on Nvidia’s new GM206 GPU. The GM206 GPU is a midrange part that’s essentially exactly half a GM204. It retains the same architectural layout and resource allocation as that GPU, with 128 CUDA cores per Streaming Multiprocessor (SM). If you’ve read our previous Maxwell coverage, than GM206 has few surprises lurking inside it. The chip may be smaller, with just 1024 cores instead of GTX 980’s 2048, but the chip’s internal caches and resources are distributed in the same ratios.
[SIZE=4]The GTX 960: Mini Maxwell[/SIZE]
One aspect of the card that’s likely to raise eyebrows is its reliance on a 128-bit main memory bus. While the GTX 960 uses fairly fast GDDR5, clocked at 1750MHz (7Gbps effective datarate), the main memory bus is tiny by modern standards.
http://www.extremetech.com/wp-content/uploads/2015/01/GTX960-Performance-640x353.jpg
A quick walk down memory lane shows just how odd it is for a GPU launching at the $199 price point to have a memory bus this narrow. The GeForce 8800 GT, which launched in 2007 at $249 had a 256-bit memory bus. All of Nvidia’s cards in this segment have historically used at least a 192-bit bus, with most opting for 256 bits and up.
This smaller bus fits with Nvidia’s general philosophy with Maxwell, which was to emphasize performance per watt and high overall efficiency over simply building a larger, more complicated chip. In fact, going solely on the numbers, the only thing extraordinary about the GTX 960 is that Nvidia is claiming it can hang with other $200 cards despite a limited memory bus and relatively few cores.
Nvidia claims the card is so efficient, it’s actually published “effective” memory bandwidth figures that inflate the card’s actual performance in an attempt to give an apples-to-apples comparison. The company is claiming that this represents an “effective” data rate as a way of communicating just how efficient the memory bus on the GTX 960 actually is. We’ll examine this in our review, but the disparity between the GTX 960 and its competition (at least on paper) is fairly significant. AMD’s R9 285, the most potent and applicable comparison card, at least as far as price is concerned, is has 176GB/s of memory bandwidth, compared to 112GB/s for the Maxwell GTX 960.
http://www.extremetech.com/wp-content/uploads/2015/01/GTX960-LiftOff-640x472.jpg
That disparity shows up when we consider the GTX 960’s primary competition, at least on paper. AMD has multiple graphics cards clustered around the $200 price point, but the obvious card to compare against is the R9 285. That GPU is based on AMD’s “Tonga’ class hardware,” and it’s the latest refresh of the GCN architecture available on the market.
[SIZE=4]Test setup and performance[/SIZE]
We evaluated the Asus Strix GTX 960 and compared it against Nvidia’s older GTX 760 and the AMD R9 285. Asus ships the Strix in a factory overclocked configuration — the GPU clock is pegged at 1291MHz base/1317MHz boost, while the memory clock is 1800MHz, up from Nvidia’s set spec of 1750MHz. In order to test the GTX 960 at proper stock speed, we underclocked the Asus Strix as well as testing it in its shipping configuration.
Unfortunately, the lowest available clock speed on the GPU core was 1200MHz, which is still roughly 4% faster than the 1178MHz Nvidia specifies as the clock on the base model. Thus, even our “Stock” clocked tests are running slightly faster than the actual shipped cards will be.
http://www.extremetech.com/wp-content/uploads/2015/01/Asus-Strix-640x372.jpg
Because vendors tend to ship a great many GPU SKUs at subtly different clocks, it’s important to capture as close to baseline as we can — but readers should be advised that many of the GTX 960 cards now shipping are clocked significantly higher than stock speed, thanks to excellent overclocking potential. We tested all cards using an Asus Z97-Deluxe motherboard and an Intel Core i5-4460 quad-core CPU with 8GB of DDR3-2133. Windows 8.1 was used, with all patches and updates installed. AMD’s latest Catalyst Omega drivers (14.12) were installed, along with Nvidia’s Forceware 347.25.
We’ve changed some of the in-game settings we use for testing overall performance to better suit the $200 price point and put a focus on higher performance as opposed to simply throwing the proverbial kitchen sink at the various cards.
AMD GPUs were tested using D3D and Mantle, where appropriate, and in some cases we evaluated multiple detail levels to test whether or not the smaller 128-bit pipe on the GTX 960 would prove a liability. We confined our testing to 1080p at the moment — the question of whether or not 4K is possible on midrange card is an interesting one, but a question for another time.
[SIZE=4]Civilization: Beyond Earth[/SIZE]
Civilization: Beyond Earth is the latest world-building game from Firaxis and a sequel to the popular Civilization V. The game supports both Mantle and Direct3D, and includes a built-in benchmark (run from the Steam command line via the “-benchmark LateGameView” command.) We tested the game at High Quality with 4x MSAA enabled, though the application identifies and applies three different types of MSAA depending on the GPU. The GTX 960 uses traditional MSAA, the GTX 760 uses 8x CSAA/4x MSAA, and the Radeon R9 285 uses 8x EQAA, 4x MSAA. There was no way to configure all three cards to use the same type of antialiasing; the specific method is auto-detected by the game engine.
http://www.extremetech.com/wp-content/uploads/2015/01/CBE.png
Civilization: Beyond Earth sets the stage for a major jump in overall performance. All of the cards are playable, but the Asus Strix 960 is fully 53% faster than the old GTX 760. AMD’s R9 285 is slightly faster when using Mantle and ties with the Asus Strix GTX 960 at its standard speed, while slightly beating our stock-clocked card.
[SIZE=4]BioShock Infinite[/SIZE]
BioShock Infinite was tested using that game’s Ultra settings with the Alternative Depth of Field option using the built-in benchmark option.
http://www.extremetech.com/wp-content/uploads/2015/01/BioShock.png
The R9 285 is again just barely faster than the Asus Strix GTX 960, with a slightly larger margin over the stock-clocked GTX 960. Nvidia’s new GPU hangs remarkably well for a card with such a small memory bus, but neither BioShock Infinite nor Civilization: Beyond Earth are particularly stressful on the GPU. Let’s take a look at what happens when we move to cutting-edge titles that put more emphasis on graphics fidelity.
[SIZE=4]Metro Last light[/SIZE]
We tested Metro Last Light in High Quality with 16x anisotropic filtering and normal tessellation. We ran this test with supersampled antialiasing both enabled and disabled — SSAA puts a huge load on a GPU and can skew results that favor one card over another. On the other hand, if the 128-bit bus on the GTX 960 is a real bottleneck, we should see that when we engage SSAA.
http://www.extremetech.com/wp-content/uploads/2015/01/MetroLastLight.png
Interestingly enough, we see performance trend in the opposite direction than expected inMetro Last Light. One would expect the GTX 960, with its 128-bit memory bus, to choke and die when SSAA is enabled. Instead, it’s the R9 285 that actually loses more performance, dropping from essentially tied with the Asus Strix GTX 960 to just slightly slower. The stock-clocked version of the card shows similar characteristics.
[SIZE=4]Total War: Rome II[/SIZE]
Total War: Rome II is the sequel to the earlier Total War: Rome title. It’s fairly demanding on modern cards, particularly at the highest detail levels. We tested Rome II in with both the “Ultra” preset and the “Extreme” preset. We tweaked both slightly — our Ultra test used antialiasing and we enabled 16x anisotropic filtering, while our “Extreme” test added high depth of field, SSAO, and vegetation alpha. These features tend to stress memory bandwidth and computation horsepower, so if the GTX 960 can’t handle a hefty load these are the settings that will show it — even at just 1080p.
http://www.extremetech.com/wp-content/uploads/2015/01/TotalWarRome2.png
Again, the GTX 960 delivers a huge improvement over the GTX 760, while eking out wins against the R9 285. The gap is slightly smaller against the stock-clocked card, and at Extreme detail, but the fact is — the GTX 960 can hang at 1080p maximum detail.
[SIZE=4]Company of Heroes II[/SIZE]
Company of Heroes is an RTS game that’s known for putting a hefty load on GPUs, particularly at the highest detail settings. The game is often CPU-bound, and that seems to be the case for everything past the GTX 760, where the stock-clocked GTX 960, the Asus Strix, and the Radeon R9 285 all end up more or less in the same place.
http://www.extremetech.com/wp-content/uploads/2015/01/COH2.png
[SIZE=4]Dragon Age: Inquisition[/SIZE]
Finally, we’re adding Dragon Age: Inquisition as our latest test. Like Civilization: Beyond Earth, DAI supports Mantle. At the top end, the GTX 980 and the Radeon R9 290X actually perform quite similarly — so how do things look at the midrange?
We performed two separate tests in Dragon Age: Inquisition — we ran the game’s built-in benchmark, and we performed a playtest during the siege of Haven. We perform the latter after noting that the built-in benchmark suffers from texture pop-in in multiple places and doesn’t accurately model real-world game play — it tends to return frame rates that are higher than those the player actually experiences.
We tested DAI using the Ultra preset in both our run-throughs.
http://www.extremetech.com/wp-content/uploads/2015/01/DAI.png
In Dragon Age: Inquistion, we see the GTX 960 and the R9 285 tie in our play test, though the GTX 960 has the edge in the built-in benchmark. Both cards significantly outperform the GTX 760. Mantle is not a huge win for AMD here, but that may be because the game’s Direct3D implementation is already very good (AMD CPUs, not GPUs, have tended to be the major winners where Mantle is concerned).
[SIZE=4]Power consumption, GPU efficiency[/SIZE]
Finally, there’s power consumption. We tested all our cards using Metro Last Light‘s High Quality option with SSAA enabled, and measured their power draw throughout the benchmark. Power consumption at idle was measured from the Windows desktop.
http://www.extremetech.com/wp-content/uploads/2015/01/PowerConsumption.png
This is where the GTX 960 truly proves its mettle. The power consumption on the GTX 960 blows both its predecessor and the R9 285 out of the water. We can evaluate this more clearly if we divide total run-time power consumption at the wall by the number of frames per second, thereby deriving the watts of power required to generate one frame per second of game rendering.
http://www.extremetech.com/wp-content/uploads/2015/01/WattsPerFrame.png
Here, we see that the GTX 760 and Radeon R9 285 are essentially on the same level — the AMD card might draw more power than Nvidia’s old midrange offering, but it also offered significantly better performance. The GTX 960, however, remains in a class of its own. It draws more than 30% less energy per frame of Metro Last Light rendered.
[SIZE=4]The GTX 960 leads on technology, AMD continues to compete on price[/SIZE]
There are multiple ways to consider these results. From a pure price/performance ratio, AMD and Nvidia remain closely tied. The R9 285 is generally matched against the GTX 960, even leading it in several tests. The two cards have equivalent amounts of VRAM, and both have their own specific feature sets meant to appeal to end users.
When it comes to overall technology leadership, however, Nvidia is clearly winning past its rival. Power consumption is dramatically better for Nvidia vs AMD. The 128-bit bus on the GTX 960 may weaken its performance at higher resolutions, but at 1080p it’s the smallest and quietest solution available. H.265 encode and decode are also performed in hardware — another potential advantage for Nvidia’s customers.
http://www.extremetech.com/wp-content/uploads/2015/01/Asus-GTX960-Strix-640x473.jpg
As we’ve said before, it’s not unusual to see this kind of gap open up in the GPU market when one company launches a new architecture before the other. AMD had a huge lead over Nvidia when it launched the HD 5000 family, and it took Team Green more than six months to answer with a GPU of its own. Midrange competition took longer; the HD 5000 family had been on the market nearly a year before Nvidia could debut the GTX 460.
AMD, in other words, still has time to craft a counter-punch, and we’ve covered rumors of upcoming hardware that speaks to Team Red’s intent to do so — but absent any kind of launch window or GPU performance figures, we can’t say when those launches will happen. Until they do, AMD is fighting at a functional disadvantage. It competes well in price/performance, but its performance in other metrics and video encode/decode is eroding its competitive stance in other areas.
http://www.extremetech.com/gaming/198105-nvidias-gtx-960-maximum-mileage-from-miniature-maxwell