NVIDIA Geforce GTX 1070 Thread

[Pheran]

Any predictions on when AIB stock will stabilize? Fortunately I'm about to go out of town for 2 weeks so that will help ease the wait a bit.

 

[Genx87]

I figured by middle of July. But who knows. Huge pent up demand for these cards.

 

[jpiniero]

It's going to be some time, especially now that nVidia is now also throwing GP106 in the mix. I guess the only saving grace is that the iPhone 7 looks like it is going to be a bust so nVidia might get a higher percentage of the 80k wafers/month that TSMC can do right now.

 

[sze5003]

I don't get why the other aib models are so expensive for example the Asus, msi, gigabyte are $469 or so. I understand the water block ones but for something that was suppose to msrp $379 or whatever, what are they adding besides the fans to jack up the price?

 

[Termie]

There is no GTX 1070 model that costs $380. That was a fiction created by Nvidia for a press event.

The only price set by Nvidia was $450 for the Founders Edition. Every other price is totally at the discretion of board partners, and no sane board partner will price anything at the minimum suggested MSRP.

If you're looking for a tear-down on why one model or another is worth "more" than the FE, you won't find it. The fact is they can sell them for whatever they want, and they will continue to do so until there is no demand remaining for these cards.

 

[ZGR]

It sounds like the GTX 1070 does in fact only use 48 ROPs, which is something a lot of sites, Anandtech included, completely ignored. Also, how much VRAM can the 1070 use that is rated at full speed? I hope we can get to the bottom of this, like the GTX 970.

Anandtech posted NVIDIA's GTX 1070 spec on May 18th:
http://www.anandtech.com/show/10336/nvidia-posts-full-geforce-gtx-1070-specs

Previously disclosed at 6.5 TFLOPs of compute performance, we now know how NVIDIA is getting there. 15 of 20 SMs will be enabled on this part, representing 1920 CUDA cores. Clockspeeds are also slightly lower than GTX 1080, coming in at 1506MHz for the base clock and 1683MHz for the boost clock. Overall this puts GTX 1070’s rated shader/texture/geometry performance at 73% that of GTX 1080’s, and is a bit wider of a gap than it was for the comparable GTX 900 series cards.

However on the memory and ROP side of matters, the two cards will be much closer. The GTX 1070 is not shipping with any ROPs or memory controller channels disabled – GTX 970 style or otherwise – and as a result it retains GP104’s full 64 ROP backend. Overall memory bandwidth is 20% lower, however, as the GDDR5X of GTX 1080 has been replaced with standard GDDR5. Interestingly though, NVIDIA is using 8Gbps GDDR5 here, a first for any video card. This does keep the gap lower than it otherwise would have been had they used more common memory speeds (e.g. 7Gbps) so it will be interesting to see how well 8Gbps GDDR5 can keep up with the cut-down GTX 1070. 64 ROPs may find it hard to be fed, but there will also be less pressure being put on the memory subsystem by the SMs.

Is the text highlighted in bold inaccurate? I am confident the GTX 1070 does in fact have 64 ROPs on board. The question is if the 1070 can actually utilize more than 48 of them. What about VRAM partitioning? Is there the infamous crossbar that Nvidia uses on the GTX 970, 660 ti, etc?

 

[Genx87]

It would help if you posted a link that is discussing 48 ROPs.

 

[zlatan]

It sounds like the GTX 1070 does in fact only use 48 ROPs, which is something a lot of sites, Anandtech included, completely ignored. Also, how much VRAM can the 1070 use that is rated at full speed? I hope we can get to the bottom of this, like the GTX 970.

Anandtech posted NVIDIA's GTX 1070 spec on May 18th:
http://www.anandtech.com/show/10336/nvidia-posts-full-geforce-gtx-1070-specs



Is the text highlighted in bold inaccurate? I am confident the GTX 1070 does in fact have 64 ROPs on board. The question is if the 1070 can actually utilize more than 48 of them. What about VRAM partitioning? Is there the infamous crossbar that Nvidia uses on the GTX 970, 660 ti, etc?

The text is accurate. The GTX 1070 has 64 ROPs and 2 MB L2 cache. There is no special VRAM partitioning or other crossbar tricks. But it has a full GPC block disabled, so the GPU loose one setup block. With the Pascal setup configuration three setup is only able to feed 48 ROPs in these synthetic scenarios.

 

[Erenhardt]

The text is accurate. The GTX 1070 has 64 ROPs and 2 MB L2 cache. There is no special VRAM partitioning or other crossbar tricks. But it has a full GPC block disabled, so the GPU loose one setup block. With the Pascal setup configuration three setup is only able to feed 48 ROPs in these synthetic scenarios.

So it has 64ROPs but 16 of those are useless?

 

[zlatan]

So it has 64ROPs but 16 of those are useless?

Not absolutely useless, but it is very hard to use them in ROP limited scenarios.

 

[Head1985]

So it has 64ROPs but 16 of those are useless?

Yes and 1060 will be same.48Rops on paper, but 1280SP=32Rops max utilisation.

 

[ZGR]

It would help if you posted a link that is discussing 48 ROPs.

Pixel fillrate is in line with 48 ROPs. This was posted by Silverforce11 in another thread.

 

[Genx87]

So it has 64ROPs but 16 of those are useless?

If I am reading him correctly. The bottleneck becomes the shaders in these synthetic benchmarks. It has 64ROPs but in a full tilt to the max synthetic scenario the shader count is only able to push the ROPs at full tilt equivalent to 48.

Every chip as a bottleneck. Synthetic scenarios are great at finding the bottleneck.

 

[Ares202]

Does how many rops the card has matter at this point? We all know the cards performance regardless or 48 or 64 rops

Unless Nvidia specifically is trying to pull the wool over customers eyes like GTX 970 3.5gb then I don't see a problem

I don't get why the other aib models are so expensive for example the Asus, msi, gigabyte are $469 or so. I understand the water block ones but for something that was suppose to msrp $379 or whatever, what are they adding besides the fans to jack up the price?

Some have totally different boards i.e MSI with (arguably) much better components than the FE version. At least the review samples tend to OC more. That could be down to the board, how the chips are being binned, the extra 6 pin power connector, or the better heat dissipation of the custom coolers, I don't think its very clear at this stage.


Also more prestigous manufacturers like EVGA can charge more due to reputation alone

 

[Creig]

Does how many rops the card has matter at this point? We all know the cards performance regardless or 48 or 64 rops

Unless Nvidia specifically is trying to pull the wool over customers eyes like GTX 970 3.5gb then I don't see a problem

A lot of people said the same thing about the 970. "We know its performance so it doesn't matter what the specs really are". Which is nonsense. What the exact specs are and how that translates to gaming performance is very relevant.

 

[DownTheSky]

Pixel fillrate. The metric where nVidia cards always trounce AMD. And then we wonder how come nVidia cards have higher performance with less TFlops.

EDIT: Why I think AMD doesn't always balance their cards properly.
Take 7970 with 2048SP and 32ROPs then take R9 390 with 2560 SP and 64ROPs. At same clock, with 25% more SP, 390 is up to 50% faster.

 

[DownTheSky]

Yes and 1060 will be same.48Rops on paper, but 1280SP=32Rops max utilisation.

For 1280 SP they don't need more. The high clock makes the difference.

 

[Sweepr]

Edit: DICE provided PCs with Core i7-6700K and Geforce GTX 1080 to test Battlefield 1. The card manages around 52 FPS @ 4K Ultra @ Closed Alpha. I think Geforce GTX 1070 should be able to hit 60 FPS on 1440p Ultra in the final build.

www.eurogamer.net/articles/digitalfoundry-2016-hands-on-with-battlefield-1-on-pc

 

[Bacon1]

So DICE is sending PCs with Core i7-6700K and Geforce GTX 1080 to test Battlefield 1. The card manages around 52 FPS @ 4K Ultra @ Closed Alpha. I think Geforce GTX 1070 should be able to hit 60 FPS on 1440p Ultra in the final build.

www.eurogamer.net/articles/digitalfoundry-2016-hands-on-with-battlefield-1-on-pc

They didn't send out review PCs, this was an event (E3? They mentioned xbox 1 version playable at Microsoft area).

 

[Sweepr]

They didn't send out review PCs, this was an event (E3? They mentioned xbox 1 version playable at Microsoft area).

They didn't send, but this was the hardware they provided to test the game at their event:

In order to best experience the game, EA provided a room full of high-end PCs featuring the latest generation Intel i7 processors paired with brand new Nvidia GTX 1080 graphics cards and G-Sync monitors.

The single GP104 card hit >50 FPS @ 4K Ultra.

 

[Ares202]

A lot of people said the same thing about the 970. "We know its performance so it doesn't matter what the specs really are". Which is nonsense. What the exact specs are and how that translates to gaming performance is very relevant.

Are you buying a GPU or rops?

if 64 rops was all over the boxes and spec sheet then I might understand you're point a little.

 

[aufkrawall]

Has anyone run the CUDA memory bandwidth testing tool with disabled DWM?

 

[thilanliyan]

Are you buying a GPU or rops?

if 64 rops was all over the boxes and spec sheet then I might understand you're point a little.

AT and other review sites have posted 64 ROPs:
http://www.anandtech.com/show/10336/nvidia-posts-full-geforce-gtx-1070-specs

Having that posted on review sites is not a whole lot worse than posting that on the box.

 

[Sweepr]

2015/2016 was not kind to multi GPU, Geforce GTX 1070 beats Radeon RX 480 CrossFire in most titles:

http://pclab.pl/art70455-23.html

 

[Sweepr]

Availability is not a problem anymore for the 1070:

www.nowinstock.net/computers/videocards/nvidia/gtx1070

Cheapest models out of stock though.