[VC] NVIDIA GeForce GTX 1060 Specifications Leaked, Faster than RX 480

[sontin]

eg. of truth and lies please

This is easy.
AMD : Fake 150W TDP vs. real 170W.
nVidia: Real 180W TDP vs Real 180W (165W precisely).

 

[Head1985]

There are only 2 possible cases:

1) If 1060 is 128-bits/32ROPs, then its specs are exactly 50% of 1080 in all metrics.
In the chart this translates to around 92.5% ~970 performance level (185 * 0.5).

2) If 1060 is 192-bits/48ROps, then its better to compare it with 1070 (which is equal to a 1080 but with fewer shaders).

For this case we have:

1070 -> 1060
1920 cc -> 1280 cc or 33% less shaders
64ROPs -> 48ROPs or 25% less ROPs
256GB/s -> 192GB/s or 25% less bandwidth
both same clocks
We can draw the conclusion for case 2: 1060 has at worst 66% the performance of 1070 (taking into account its weakest point, shaders).

In chart it would be like 101% ~390/480 performance. (153 * 0.66)

In any case I dont see it being 15% faster than 480..

Well 1070 have only 48Rops because 1070 have only 1920SP.There are 64Rops, but 1070 can only Use 48Rops
1060 and 1070 will be pretty close in performance.. Around 30% gap.

Btw you can see it at pixel fillrate test.Also GTX1070 is most likely only 3X GPC SKU vs 4x GPC 1080.

EDIT:with 1280SP GTX1060 can only use 32Rops.

EDIT2:
Real life specs
GTX1070
1920SP
48Rops
256bit

GTX1060
1280SP
32rops
192bit

So 1070 will have 50% more SP and 50% more Rops.But only 33% more memory bandwidth.

 

[AnandThenMan]

Whatever man. You're set in your ways. I certainly hope you didn't expect anyone to just keel over and follow your opinionated advice blindly.

RS posted cold hard facts not opinions.

 

[renderstate]

[Nope]

Infraction issued for member callout.

-Rvenger

 

[iiiankiii]

Whatever man. You're set in your ways. I certainly hope you didn't expect anyone to just keel over and follow your opinionated advice blindly. People have their own minds. Yes, it's astonishing, but they do. Not everyone sees things your way, and just because it's your way, doesn't necessarily make it the "right" way for anyone else. If it's right for you, beautiful.
You are 10000000% about performance to the last cent. In which case, you will always be in AMDs corner, because all they can ever do is compete on a price/performance aspect. They are the underdog, they have to do a LOT more to earn the same money. That is plainly obvious. blow for blow, Nvidia earns more. Prices can be higher. People will buy them. They are fantastic products as are AMDs. There really isn't anything you can do to change that.

It really depends on the price tier. You're right. It isn't always cut and dry as RS seems to make it out sometimes. However, perf/$ is THE NUMBER ONE metric when you buy a GPU for most people. I understand there are outliers that demand certain things which force them to pick one GPU over the other. I can respect that.

BUT, overall, I find myself recommending AMD GPUs over Nvidia's GPUs most of the time. The price premium Nvidia demands is just a luxury that isn't necessary needed (nor wanted). Most people just want X card to run X games. It's just a means to an end. In that respect, AMD does a fine job at a lower price point.

 

[linkgoron]

RS is so concerned the 1060 is going to steal the 480 thunder that he/she is in full preemptive defense mode, coming up with all sorts of convoluted reasons for which the 480 is still a better buy even if the 1060 soundly beats it in every metric except perf/$. God forbid someone could decide to spend a little more money for a better/more well rounded product.

Truth is, NVIDIA architecture is currently more efficient than GCN and NVIDIA is executing generally better than AMD. If the current trend carries over to the 1060 there is no reason to think it won't be a more successful product than the 480.

We will know soon whether this is the case or not.

perf/$ means a lot, but I agree that RS is too hard on the perf/$ and percentages sometimes. Sometimes people just want the performance, or just want to spend X money, or just want 60FPS+ or whatever.

However, I agree with him that $ is important.
I'm betting that the 1060 will be around ~5% faster than the reference 480 in DX11 at 1080p (don't know about dx12).

If the above is true:
For 250$ - the 3gb is not worth it over the 480.
For 250$ - the 6gb is worth it over the 480.
For 300$ - I would buy a 1070 and not a 1060.

The above is obviously pending reviews of AIB 480 and reference/aib 1060.

 

[mkmitch]

RS posted cold hard facts not opinions.

Using phrases like, money spent better elsewhere and nothing special are not necessarily indicative of cold hard facts, rather they are his opinions. :wub: I get alot of info from RS but his preachy manner is irritating and discounts the value in what he says.

 

[Wall Street]

Well 1070 have only 48Rops because 1070 have only 1920SP.There are 64Rops, but 1070 can only Use 48Rops
1060 and 1070 will be pretty close in performance.. Around 30% gap.

Btw you can see it at pixel fillrate test.Also GTX1070 is most likely only 3X GPC SKU vs 4x GPC 1080.

EDIT:with 1280SP GTX1060 can only use 32Rops.

EDIT2:
Real life specs
GTX1070
1920SP
48Rops
256bit

GTX1060
1280SP
32rops
192bit

So 1070 will have 50% more SP and 50% more Rops.But only 33% more memory bandwidth.

This is not right at all. The 1070 has 64 ROPs. ROPs are tied to the memory controller and the L2 caches. If it only had 48 ROPs then it would only be able to address 6 GB of the memory - or have a GTX 970 like memory issue - that was caused by disabling ROPs. This is not the case with the GTX 1070 which has the same amount of ROPs as the 1080 but 3/4 the number of shaders.

EDIT: However, the pixel rate of the GTX 1070 can be constrained by it limited shaders. In the case of the GTX 1070, its shaders can output 60 pixels/clock (15 SMs x 4 pixels per SM) which would not be enough to saturate the 64 ROPs, but is close.

 

[Janooo]

Whatever man. You're set in your ways. I certainly hope you didn't expect anyone to just keel over and follow your opinionated advice blindly. People have their own minds. Yes, it's astonishing, but they do. Not everyone sees things your way, and just because it's your way, doesn't necessarily make it the "right" way for anyone else. If it's right for you, beautiful.
You are 10000000% about performance to the last cent. In which case, you will always be in AMDs corner, because all they can ever do is compete on a price/performance aspect. They are the underdog, they have to do a LOT more to earn the same money. That is plainly obvious. blow for blow, Nvidia earns more. Prices can be higher. People will buy them. They are fantastic products as are AMDs. There really isn't anything you can do to change that.

RS makes a point.
I do not understand what you are saying except all the jabs you through at him.
Wow...

 

[sm625]

I dont get where the $250 pricetag on a 3GB 1060 is coming from. That card is very likely to have a BoM that is 20% less than the 480. A smaller die, less power = less VRMs, only 3/4 of the memory chips, smaller pcb, cheaper thermal solution, etc. Nvidia could easily set MSRP to $199 and still make bank. Could even go $189.

 

[sze5003]

I dont get where the $250 pricetag on a 3GB 1060 is coming from. That card is very likely to have a BoM that is 20% less than the 480. A smaller die, less power = less VRMs, only 3/4 of the memory chips, smaller pcb, cheaper thermal solution, etc. Nvidia could easily set MSRP to $199 and still make bank. Could even go $189.

$199 is more like it for 3gb knowing how they price their cards, usually it's on the high side.

 

[IEC]

I dont get where the $250 pricetag on a 3GB 1060 is coming from. That card is very likely to have a BoM that is 20% less than the 480. A smaller die, less power = less VRMs, only 3/4 of the memory chips, smaller pcb, cheaper thermal solution, etc. Nvidia could easily set MSRP to $199 and still make bank. Could even go $189.

That would actually be nice. If the 6GB version comes in at $249 and competes with or beats the RX 480 we could see some price cuts or promotions which would make the $200 price point for graphics cards nicer than it has been in recent memory.

 

[Arachnotronic]

I dont get where the $250 pricetag on a 3GB 1060 is coming from. That card is very likely to have a BoM that is 20% less than the 480. A smaller die, less power = less VRMs, only 3/4 of the memory chips, smaller pcb, cheaper thermal solution, etc. Nvidia could easily set MSRP to $199 and still make bank. Could even go $189.

Why would NVIDIA charge less than what it's worth? If the GTX 1060 3GB outperforms the Radeon RX 480 4GB, then it's going to be priced higher.

 

[ocre]

Checked out this thread, utterly amassing. Obvious we have a large amount of people who are unwavering in their AMD support but threads like these might suggest.... blindly.

As to be specific, there is more not known when it comes to the 1060 than we do know. Yet the huge rush to push down this up and coming Nvidia card while promoting the 480, it's shocking.

These days, there are not too many technical discussion of great caliper any more. I accepted that long ago. But....this???

Having so little info, a cheap looking graph that is far from verified and no info at all in regards to price, but the pitchforks are out and the swords are swinging. We know nothing about this card, nothing concrete and nothing for sure yet it's the same Ole Nvidia attacks from fluff marketing to their prices, yadi-yadi-ya. Same stuffing in every thread.

I personally have no idea where the 1060 will fall. Where it will perform nor what it will cost. If I assume the info leaked had some basis in reality, then it seems like it could be much more efficient than the 480. Some of the more interesting things, most technically interesting, it is the bus. If true, Nvidia went with a unique route for the gp106. Extremely interesting to me. Is this a gp106? Could it be a gp104 further cut instead? If this is their full flat gp106, I th I nk it is very interesting for Nvidia to go proportionately wider on their smaller chip. Efficiency trade off is just one thing to keep in mind with an increase bus width, other trade offs as well, such as complexity and yields. It's not that 192 is large, put it is very much so compared to the layout of the gp104. It also makes me wonder, if this is truly the gp106 then what happens with the smaller chips? Are we gonna see gp107/108 divide out of the gp106?

I think maxwell had an interesting and tight nit flow from gddr-bus-rops..it was ultra effective and efficient. The layout was designed in harmony, the data flow was built in the bounds of this configuration built entirely on effectiveness and efficiency. In such a way, that the 980 would fair no better with a 384 bit bus than a 256, unless the Rop, memory, and cores increased proportionately with this larger bus.

So basically, what I am saying is 1280 cores would be evenly split down, a gp104 cut in half. The bus, 128 would be suited and in line with the gp104. Not only would it fit the formula, it would also have memory available in 4 or 8gb..not odd ball 3 or 6gb.
Having a 192bit bus, to be truly useful...the gp106 would have to be designed different, have a completely different data stride and structural layout than the gp104. If pascal is mostly just maxwell, the added bus over 128 would not be very useful. In the maxwell design, the 970 cut..even though it technically had 256bit bus and all its rops, there was little point at all to them, they were lame and pretty much useless. Pascal could be different from maxwell. But I still have to question how this 192 bus is tied to rops and then cores.

It could be the case that this layout will be more effective, giving the 1060 more performance per Cuda core but at the expense of an increase in power consumption. If we look at the supposed tdp, 130watt is higher...a bit higher than I would expect. So, perhaps this larger bus is a trade off to gain performance on the gp106 line. It's a very interesting thing to consider. It's like Nvidia is turbo charging their x06 series much like they did back when the started turbo charging the x04 lines. If this is the case, the 106 may not be half of the 104s anymore. More like 75%

Anyway, if we think about perf per watt when looking at maxwell. It was a pretty good estimate for performance. While not exact, it was enough to generalize performance. Pascal performance is known in regards to its performance per watt. If the 1060 is 130warts, no matter what the card is (106/104), it should be faster than half of a 1080. I can't see efficiency dropping that much to where 120watts is only half a 1080 in performance. That would be a massive drop in efficiency. Just the rumor tdp alone, I would estimate the 1060 well over a 1080. It doesn't seem logical to have a severe nose dive like that. It will perform somewhere in line with the tdp. I doubt it will be as efficient as the 1080, but it should not be way too far off.

 

[Wall Street]

So basically, what I am saying is 1280 cores would be evenly split down, a gp104 cut in half. The bus, 128 would be suited and in line with the gp104. Not only would it fit the formula, it would also have memory available in 4 or 8gb..not odd ball 3 or 6gb.
Having a 192bit bus, to be truly useful...the gp106 would have to be designed different, have a completely different data stride and structural layout than the gp104. If pascal is mostly just maxwell, the added bus over 128 would not be very useful. In the maxwell design, the 970 cut..even though it technically had 256bit bus and all its rops, there was little point at all to them, they were lame and pretty much useless. Pascal could be different from maxwell. But I still have to question how this 192 bus is tied to rops and then cores.

It could be the case that this layout will be more effective, giving the 1060 more performance per Cuda core but at the expense of an increase in power consumption. If we look at the supposed tdp, 130watt is higher...a bit higher than I would expect. So, perhaps this larger bus is a trade off to gain performance on the gp106 line. It's a very interesting thing to consider. It's like Nvidia is turbo charging their x06 series much like they did back when the started turbo charging the x04 lines. If this is the case, the 106 may not be half of the 104s anymore. More like 75%

Anyway, if we think about perf per watt when looking at maxwell. It was a pretty good estimate for performance. While not exact, it was enough to generalize performance. Pascal performance is known in regards to its performance per watt. If the 1060 is 130warts, no matter what the card is (106/104), it should be faster than half of a 1080. I can't see efficiency dropping that much to where 120watts is only half a 1080 in performance. That would be a massive drop in efficiency. Just the rumor tdp alone, I would estimate the 1060 well over a 1080. It doesn't seem logical to have a severe nose dive like that. It will perform somewhere in line with the tdp. I doubt it will be as efficient as the 1080, but it should not be way too far off.

I think that there is some logic to increasing the memory bus width from 128-bit for the 1060. A 128-bit 8 Ghz bus would only be 14% faster than the GTX 960's 7 Ghz memory bus. That is too slow in 2016 when any card in the $200+ range should be able to max out 1080p. With a 128-bit bus, you could run into bandwidth problems with MSAA. Note that some effects, like MSAA, can be bandwidth intensive but aren't driven by the shaders. While a 192-bit bus would have excess bandwidth, I would worry that 128-bits would be limited. No matter how many shaders or how high a core clock, I doubt that even nVidia can get a 128-bit card to be faster than a GTX 980 (especially since the 980 already had pretty good delta color compression).

The number of ROPs is tied to the memory bus, so a 192-bit card would have 48 ROPs.

 

[3DVagabond]

I think to decide between a 480 and a 1060 i would choose the one with better perf/$. Both consume tiny amounts of power (less than a 970), you wont see a difference in electricity bill with a 10/20watts diference. Also exepect 480s to be as cool with aftermarket coolers..

I would strongly consider which one is better in newer games. Especially DX12 and Vulkan. I doubt the 15% nVidia is talking is in DX12.

 

[krumme]

How much power does ram consume on a card like 1060 with 6gb vs full consumption in games? How much for 3gb?

 

[Head1985]

This is not right at all. The 1070 has 64 ROPs. ROPs are tied to the memory controller and the L2 caches. If it only had 48 ROPs then it would only be able to address 6 GB of the memory - or have a GTX 970 like memory issue - that was caused by disabling ROPs. This is not the case with the GTX 1070 which has the same amount of ROPs as the 1080 but 3/4 the number of shaders.

EDIT: However, the pixel rate of the GTX 1070 can be constrained by it limited shaders. In the case of the GTX 1070, its shaders can output 60 pixels/clock (15 SMs x 4 pixels per SM) which would not be enough to saturate the 64 ROPs, but is close.

Like i said before.1070 cant use 64rops because it only have 1920SP.It can use only 48Rops with 1920SP.

Pixel fillrate is much lower than GTX1080 but both have 64rops and should have same results..1080 have 33% more pixel fillrate.48rops vs 64rops=33%more also 1920SP vs 2560SP=33%more.

And BTW 1060 can only USE 32Rops with 1280SP.

 

[3DVagabond]

Using phrases like, money spent better elsewhere and nothing special are not necessarily indicative of cold hard facts, rather they are his opinions. :wub: I get alot of info from RS but his preachy manner is irritating and discounts the value in what he says.

I think you are attributing something to him that isn't there. He's being analytical. Nothing more or less.

You might not notice but there are plenty of times where he recommends nVidia over AMD. That's because it's what his analysis tells him. He was a big 970 and 980 ti proponent when they were first released. I used to disagree with him because I factored in the likelihood that AMD's performance would improve over time relative to what it was then. Although neither of us accuse the other of being biased. We simply disagree.

 

[3DVagabond]

I dont get where the $250 pricetag on a 3GB 1060 is coming from. That card is very likely to have a BoM that is 20% less than the 480. A smaller die, less power = less VRMs, only 3/4 of the memory chips, smaller pcb, cheaper thermal solution, etc. Nvidia could easily set MSRP to $199 and still make bank. Could even go $189.

We don't know what AMD pays GloFo vs. what nVidia pays TSMC. There's also the WSA to factor in. Just because it's economically advantageous for AMD to sell the 480 for $200 doesn't mean it would be for nVidia as well.

 

[Silverforce11]

Like i said before.1070 cant use 64rops because it only have 1920SP.It can use only 48Rops with 1920SP.

Pixel fillrate is much lower than GTX1080 but both have 64rops and should have same results..1080 have 33% more pixel fillrate.48rops vs 64rops=33%more also 1920SP vs 2560SP=33%more.

And BTW 1060 can only USE 32Rops with 1280SP.

I thought it was suspicious that the 1070 diagram show a 1/4 cut, a cut at the GPC means ROPs are gone as well but they list it full.

Looks like another round of 970 specs, that fillrate test proves it's only got 48 ROPs usable.

Did anyone do a memory bandwidth test?

Because NV's MC & ROPs are linked, if ROPs aren't usable, then the full 256 bit bus isn't usable nor will the full 8GB vram be usable at full speed..

 

[SolMiester]

WTH?
The GTX750TI was the competence to 260/260X/360, NOT 370/270/270X, AMD launched the R7 265 to better fight the 750TI at its price point, but AMD give up fighting the 750TI on 300 series on performance and just decided to fight it on price with the 360.

GTX950 is the competence to R7 370, NOT R9 380s...


AMD did have a good product on the R9 380 vs GTX960, but thats about it, i whould have picked a GTX750TI over a 360/260, and GTX950 over the GCN 1.0 R7 370 ANY DAY, those NV cards where just better, specially on DX11 with slowers CPUs.

AND were they benched with budget CPUs as the will probably be paired with by those buying the budget cards, or i5s & i7s?

 

[Keysplayr]

RS posted cold hard facts not opinions.

Opinions

 

[Silverforce11]

I can understand this line of reasoning AFTER benchmarks indicate this is the case, but not before and certainly not with such surety.

Pretty certain GP106 will be more efficient, ie, perf/w than Polaris 10.

There's just no way for a cut down half GP104, GP106 to be a 150W GPU. It'll be close to 100-120W, and if it has performance around 970/980, it will be ~35% more perf/w.

However, in DX12 & Vulkan games, that changes obviously.

Doom's getting a Vulkan patch in a few days I heard, that will be interesting. Coupled with TW:Warhammer's DX12, we have a few blockbuster games on next-gen API by the time the 1060 lands. More to come in the next few months.

Let's face it, whichever card runs Battlefield 1 (especially this one), Deus Ex and Watch Dogs 2 faster is the winner simply due to how big these titles/PR are.

 

[JDG1980]

There's just no way for a cut down half GP104, GP106 to be a 150W GPU. It'll be close to 100-120W, and if it has performance around 970/980, it will be ~35% more perf/w.

TPU lists a peak gaming power measurement of 118W for GTX 960, compared to 184W for GTX 980. The GTX 1080 matches the 980 exactly in power usage.

Does this mean we should expect the same figure (roughly 120W) for GTX 1060? Maybe. The GTX 960 ran at the same clock rate as the 980, and leaks indicate that the 1060 will likewise share the same clock rate used by the 1080. But there's one important difference: while GM206 was half of a GM204 in pretty much every respect, GP106 has a memory bus 75% as wide as that of the GP104, and with a clock rate just as high as the GTX 1070. The memory controller is often one of the more power-hungry aspects of a video card. This means GTX 1060 might go in closer to 130W peak gaming than 120W. I suppose we'll find out soon enough.

GTX 1060 FE will definitely beat reference RX 480 in perf/watt, based on what we've seen so far. Depending on how AMD fixes the power management issues, it might be closer than most people are assuming; undervolting could get RX 480 down from ~160W to 130-140W. (Pushing too much voltage has been AMD's Achilles heel for a long time.) Most people will eventually be buying AIB cards instead of reference for both RX 480 and GTX 1060, so it'll be interesting to see how this competition shakes out in both raw performance and perf/watt.

A hypothetical P10 card with GDDR5X, better binning, and more precise voltage settings could match GTX 1060's perf/watt or perhaps even beat it. Higher memory clocks mean higher power consumption (this is one reason HBM is so efficient: it uses a very wide and slow memory bus). Switching from 8GB/s GDDR5 to 10GB/s GDDR5X would let AMD drop the memory clock from 2000 MHz to 1250 MHz while simultaneously increasing bandwidth by 25%. That could make a substantial difference.