Single page Print

Nvidia's GeForce GTX 1060 graphics card reviewed


Double trouble
— 11:54 AM on October 25, 2016

Nvidia is on quite the roll this year. The GeForce GTX 1070 and GeForce GTX 1080 remain the uncontested performance champions of the high-end graphics card market, thanks in part to AMD's more mainstream ambitions for its Polaris-powered graphics cards. If that dominance wasn't enough, Nvidia did itself one better and advanced its single-GPU performance lead with the GP102 chip in the Pascal Titan X. Y'know, just because.

Of course, the green team didn't ignore the average Joe while it was busy pushing the limits of graphics performance. Back in July, Nvidia introduced the GeForce GTX 1060, its response to the Radeon RX 480 8GB. The $250 GTX 1060 was the first card to play host to a more wallet-friendly Pascal GPU: GP106. The GTX 1060 6GB, as we now know it, immediately went to work against the hard-to-get RX 480 with a slew of readily-available aftermarket cards that stickered near Nvidia's $250 suggested price tag. The mainstream onslaught didn't stop there, however. A couple weeks later, Team Green took the wraps off a GTX 1060 with 3GB of RAM that rang in at $200.


The GP106 GPU.

While the GTX 1060 3GB's name might imply a simple halving of its RAM versus its bigger brother, there's more going on under the hood of that card than its innocuous name might suggest. The GTX 1060 3GB sustained some cuts to its graphics-processing resources to hit its price target. Nvidia disabled one of the card's shader multiprocessor (SM) blocks, dropping the GTX 1060 3GB's resource allocation to 1152 stream processors and 72 texture units. Contrast that approach with AMD's Radeon RX 480 4GB, whose only difference from its 8GB cousin is that 4GB of missing RAM. Here's how the two "GTX 1060s" compare on paper, in convenient tabular form: 

  Base
clock
(MHz)
Boost
clock
(MHz)
ROP
pixels/
clock
Texels
filtered/
clock
(int8/
fp16)
SP
TFLOPs
Stream
pro-
cessors
Memory
path
(bits)
Memory
transfer
rate
(Gbps)
Memory
bandwidth
(GB/s)
Peak
power
draw
RX 470 926 1206 32 128/64 4.9 2048 256 6.6 211 120W
RX 480 1120 1266 32 144/72 5.8 2304 256 7 224 150W
GTX 960 1126 1178 32 64/64 2.4 1024 128 7.01 112 120W
GTX 970 1050 1178 56 104/104 3.9 2048 256 7.0 224 145W
GTX 1060 3GB 1506 1708 48 72/72 3.9 1152 192 8.0 192 120W
GTX 1060 6GB 1506 1708 48 80/80 4.4 1280 192 8.0 192 120W
GTX 1070 1506 1683 64 120/120 7.0 1920 256 8.1 259 150W


A block diagram of the GP106 GPU. Source: Nvidia

Nvidia has played this kind of name game with its cards before. Recall that the GeForce GTX 460 came in 768MB and 1GB flavors. Despite the identical name on the box, the lesser GTX 460 was down eight ROPs and had a narrower path to memory than its better-endowed counterpart. We complained about that false equivalency then, and we're complaining about it now. AMD isn't ashamed of putting a smaller number on its cut-down Polaris 10 card, the Radeon RX 470, and we don't think calling the GTX 1060 3GB... well, anything other than a GTX 1060 would have hurt its perception in the marketplace that much.

Thanks to that questionable naming scheme, the uninformed builder picking one of these cards off the shelf probably won't notice that more is missing from the GTX 1060 3GB than 3GB of RAM—assuming those specs are clearly spelled out on the box at all. In the case of the EVGA cards we have on hand, we found no mention of stream processor or texturing unit counts on the cards' packaging. We think that Nvidia's board partners should be more upfront about what buyers are getting if there's as substantial a difference between cards as there is between these two, even if that information is a little arcane.

  Peak pixel
fill rate
(Gpixels/s)
Peak
bilinear
filtering
int8/fp16
(Gtexels/s)
Peak
shader
arithmetic
rate
(tflops)
Peak
rasterization
rate
(Gtris/s)
Memory
bandwidth
(GB/s)
GeForce GTX 1060 3GB 82 123/123 3.9 3.4 192
GeForce GTX 1060 6GB 82 137/137 4.4 3.4 192
GeForce GTX 960 38 75/75 2.4 2.5 112
GeForce GTX 970 61 130/130 3.9 4.7 224
GeForce GTX 980 78 156/156 5.3 5.0 224
GeForce GTX 1070 108 202/202 7.0 5.0 259

Some quick math shows that a full-fat GP106 chip has slightly more raw pixel throughput and slightly less texturing muscle than the GTX 980. It's also slightly less capable than GM204 in sheer number-crunching power and memory bandwidth, although the Pascal architecture's improved delta-color-compression facility might help make up some of that gap. Of course, both GTX 1060s utterly wipe the floor with the GM206 chip that powered the GTX 960 and the GTX 950. It's a testament to the power of Pascal that we're comparing the $250 GTX 1060 6GB to cards that used to cost $350 to $500-ish.

In another potentially controversial move, Nvidia removed SLI support from both GTX 1060s. At least one set of SLI fingers has been available on every GeForce card in recent memory except for the GTX 750 Ti and below, so this move marks a new era for the spec sheets of budget-friendly GeForces. Some DirectX 12 multi-adapter modes might let gamers harness multiple GTX 1060s in the future, but the option is no longer available in DirectX 11 titles, full stop.

In recent years, we've suggested that gamers get the best single graphics card they can afford for the most consistent and smoothest possible performance in games, so we're not bothered much by this move. Folks willing to tolerate SLI's inconsistent performance scaling and potential frame-pacing issues will be disappointed by this omission, however, especially considering that the price for a pair of GTX 1060 6GB cards ends up somewhere in between a GTX 1070 and a GTX 1080. If the GTX 1060 6GB delivers on Nvidia's promise of GTX 980-class performance, a pair of those cards might have approached a GTX 1080 in raw speed, so it's not hard to imagine why the green team made this choice. Those brave folks willing to pair multiple budget graphics cards for a potential performance boost will need to stick with Radeons for now.

Now that we have the lay of the land for the GTX 1060, let's see how EVGA has chosen to put the chip to work on a pair of its graphics cards.