Geforce GTX 1060 Thread: faster than RX 480, 120W, $249

[boozzer]

haha, exactly, no benchmarks are out yet, so why are some of you guys supporting the claim in the thread title?

hell, 2 threads. 3 if the 3rd one didn't get axed.

 

[Arachnotronic]

Isn't SLI and Crossfire soon to be pointless because DX12 allows you to run multiple gpus mix and match?

I doubt that game developers are going to put in the effort required to make that happen.

 

[96Firebird]

Sad, no benchmarks even out yet, and AMD fans are already pre-preemptively trying to discredit them.

Maybe they learned from their mistake of overhyping the RX 480, and have now gone the complete opposite direction for the GTX 1060? Set themselves up for disappointment, only to be pleasantly surprised at launch!

I wish that were true, but we all know the real reason... :\

 

[NTMBK]

Isn't SLI and Crossfire soon to be pointless because DX12 allows you to run multiple gpus mix and match?

In all 4 games that support that, sure. DirectX 12 implementations are hard enough as it is, without trying to support multi-GPU- let alone any random combination of GPUs. I think most developers will target the 99% case, i.e. single GPU, and not bother with multi-GPU support.

 

[PhonakV30]

Sad, no benchmarks even out yet, and AMD fans are already pre-preemptively trying to discredit them.

Ofc , they always do , what did you expect ? Nvidia Fans will do same.nothing more.

 

[NTMBK]

The problem is no card not even the 1080 or the 1080Ti will be able to max all games, so just buying a single card is not going to always work out too. I believe it should be left to the consumers what they prefer because in the past sometimes it was a wiser option.

In the past it was never a wise option, just people didn't have good frametime measurements to show the stuttering.

 

[Jaydip]

In the past it was never a wise option, just people didn't have good frametime measurements to show the stuttering.

Has there been any recent frame-time measurements to support this? if their vram is not gimped it shouldn't' be bad.

 

[Azix]

Saw this on Baidu:

July 10 release!
Units Shading: 1280
TMUs: 80
ROPs: 48 ----------------
Pixel Fillrate: 72.3 GPixel/S

In comparison:
GTX980 pixel fillrate: 72.1 GPixel/s
RX480 pixel fillrate: 35.8 GPixel / s

Performance is equal to the public (reference) version of GTX980

whats the significance? 980 is not twice as fast as a 480. Useless specs. I'm still sticking to tflops as the more important figure and what it has shown is that pascal needs more to beat maxwell. The 1060 without boost has significantly less than stock 980. A 1060 boosting to 1800Mhz still has less, but close enough to somewhat equal a 980 at stock non-boost clocks. 980 base clock perf without OC on the 1060 is possible with high boost clock but pushing it.

Actually I expect reviews will show it at that boost clock and more, equaling the 980. Probably not so much if they warm up the GPU first, like with the reference 1080s.

 

[AtenRa]

Saw this on Baidu:

July 10 release!
Units Shading: 1280
TMUs: 80
ROPs: 48 ----------------
Pixel Fillrate: 72.3 GPixel/S

In comparison:
GTX980 pixel fillrate: 72.1 GPixel/s
RX480 pixel fillrate: 35.8 GPixel / s

Performance is equal to the public (reference) version of GTX980

For it to have 48 ROPs, either GTX 1060 is a cut down GP104

or

GP106 has a new layout of 3x GPCs with 8x ROPS and its not half the GP104 but a new design.
Then If it has 3X GPCs, for it to have 1280 Cores, then Each TPC design will be different than GP104.

 

[Arachnotronic]

For it to have 48 ROPs, either GTX 1060 is a cut down GP104

or

GP106 has a new layout of 3x GPCs with 8x ROPS and its not half the GP104 but a new design.

Is it really so far-fetched to think that NVIDIA has the resources to develop a chip suitable for a given target market? I don't think so.

 

[redzo]

In the past it was never a wise option, just people didn't have good frametime measurements to show the stuttering.

True.

I would add.

Multi-gpu gaming is dead. Nvidia wants it dead. Investors are too cheap and they will never support multi-gpu: no precious development time(money) spent at optimizing engines for multi-gpu. Historically speaking, AMD is incapable to gather enough developers to make a possible 2*480 config relevant for gaming. Just get a damn 1070 for the same price and deal with it! Too many drawbacks with a 2*480 config. Unless you are cherry picking a particular game or a very specific use case, it's a dead end.

For it to have 48 ROPs, either GTX 1060 is a cut down GP104

or

GP106 has a new layout of 3x GPCs with 8x ROPS and its not half the GP104 but a new design.
Then If it has 3X GPCs, for it to have 1280 Cores, then Each TPC design will be different than GP104.

Probably a new design. It looks like it's cut down too much

 

[Azix]

True.

I would add.

Multi-gpu gaming is dead. Nvidia wants it dead. Investors are too cheap and they will never support multi-gpu: no precious development time(money) spent at optimizing engines for multi-gpu. Historically speaking, AMD is incapable to gather enough developers to make a possible 2*480 config relevant for gaming. Just get a damn 1070 for the same price and deal with it! Too many drawbacks with a 2*480 config. Unless you are cherry picking a particular game or a very specific use case, it's a dead end.

$350 for a 970 4Gb. 2 x 970 = omg buy it. so great. wonderful. much value.

$200-$250 for a 480 4GB/8GB. Can get up to $700 performance. Oh no, multiGPU is dead. Not worth. DoooOOoon't.

:thumbsdown:

the situation with multiple GPUs has not changed, oh except at least one company is on record wanting to improve it. The other is being a stick in the mud as usual, but they aren't going to cut sli. selling multiple GPUs to the same person is desirable. Investing in getting things working well for that to become more common is sensible. But I am guessing they will let AMD push it in dx12 then take advantage of that

 

[MrTeal]

For it to have 48 ROPs, either GTX 1060 is a cut down GP104

or

GP106 has a new layout of 3x GPCs with 8x ROPS and its not half the GP104 but a new design.
Then If it has 3X GPCs, for it to have 1280 Cores, then Each TPC design will be different than GP104.

Why could each TPC GPC not have 3x8 ROPs linked to it? Each GPC in GP104 has five TPC's linked to two 8 ROP banks according to the block diagrams. Three banks of ROPs for 5 TPC's doesn't seem any stranger than two for 5 TPCs.

 

[redzo]

$350 for a 970 4Gb. 2 x 970 = omg buy it. so great. wonderful. much value.

$250 for a 480 8GB. Can get up to $700 performance. Oh no, multiGPU is dead. Not worth. DoooOOoon't.

:thumbsdown:

I was never a proponent of multi-gpu. Not only that I'm being consistent, but times have changed. Multi-gpu lays now in the hands of developers, at the mercy of a development business decision.

 

[AtenRa]

Is it really so far-fetched to think that NVIDIA has the resources to develop a chip suitable for a given target market? I don't think so.

Im not implying something like that,

On the other hand, if it has 80x TMUs, that implies 2x GPCs with 10x TPCs and that is exactly half the GP104.

 

[Rvenger]

Fixed that for you. Next time try to stick to the facts instead of going personal on every instance.

Next time you alter a quote you will be gone for 30 days.

-Rvenger

 

[AtenRa]

Why could each TPC not have 3x8 ROPs linked to it? Each GPC in GP104 has five TPC's linked to two 8 ROP banks according to the block diagrams. Three banks of ROPs for 5 TPC's doesn't seem any stranger than two for 5 TPCs.

In GP104, there are 4x GPCs and each GPC has 5x TPCs and each TPC has 64x Cores.
There are 16x ROPs per each GPC, ROPs are decoupled from the Memory controllers.

Edit :

If there are 3x 8 ROPs per GPCs then its a new design vs GP104.

Edit: Sorry 5x TPCs per GPC.

 

[MrTeal]

In GP104, there are 4x GPCs and each GPC has 10x TPCs and each TPC has 64x Cores.
There are 16x ROPs per each GPC, ROPs are decoupled from the Memory controllers.

Edit :

If there are 3x 8 ROPs per GPCs then its a new design vs GP104.

Sorry, wrong acronym there, I meant GPC.

I thought GP104 had five TPC's per GPC, but the basic idea doesn't change. The block diagram shows the ROPs outside the GPCs, and while the two banks of 8 ROPs are likely tied to one GPC, it seems a lot less of a design change to increase that to three banks of 8 ROPs per GPC than it does to alter the number of TPCs in a GPC, the number of SM's in a TPC, or the number of CC's in a SM.

 

[Wall Street]

There are 16x ROPs per each GPC, ROPs are decoupled from the Memory controllers.

This is not true. ROPs are decoupled from the GPCs and tied to the memory controller and L2 cache.

 

[MrTeal]

This is not true. ROPs are decoupled from the GPCs and tied to the memory controller and L2 cache.

That would make a lot more sense. One bank of 8 ROPs per 32-bit MC. Since GP106 is rumoured to be a 192-bit design, you would expect 48 ROPs.

 

[AtenRa]

Sorry, wrong acronym there, I meant GPC.

I thought GP104 had five TPC's per GPC, but the basic idea doesn't change. The block diagram shows the ROPs outside the GPCs, and while the two banks of 8 ROPs are likely tied to one GPC, it seems a lot less of a design change to increase that to three banks of 8 ROPs per GPC than it does to alter the number of TPCs in a GPC, the number of SM's in a TPC, or the number of CC's in a SM.

Well the thing is, if you only keep 2x GPCs with 5x TPCs each, you only have 10x Polymorph engines, when GM204 (GTX 980) has 16x. That will severely negatively impact the GTX 1060 tessellation performance and it will have a very though time against GTX 980.

 

[AtenRa]

This is not true. ROPs are decoupled from the GPCs and tied to the memory controller and L2 cache.

Do you have any documentation for this ???

 

[MrTeal]

Well the thing is, if you only keep 2x GPCs with 5x TPCs each, you only have 10x Polymorph engines, when GM204 (GTX 980) has 16x. That will severely negatively impact the GTX 1060 tessellation performance and it will have a very though time against GTX 980.

Yeah, but GP104 only has 20 Polymorph engines (15 in the 1070) which is less than the 24 in GM200, yet it has no trouble in high tesselation situations vs a Titan X. The 1060 having 2/3rds of the PEs of the 1070 doesn't seem like it should be a massive issue when GM204 has the same ratio relative to GM200 and the 1070 doesn't seem to be limited in tessellation relative to the Titan X.

 

[AtenRa]

Yeah, but GP104 only has 20 Polymorph engines (15 in the 1070) which is less than the 24 in GM200, yet it has no trouble in high tesselation situations vs a Titan X. The 1060 having 2/3rds of the PEs of the 1070 doesn't seem like it should be a massive issue when GM204 has the same ratio relative to GM200 and the 1070 doesn't seem to be limited in tessellation relative to the Titan X.

Do we have any review with the GTX 1070 vs Titan X and games like Witcher 3 with Hairworks enabled ?? Because i have my doubts the GTX 1070 will be faster than Titan X in heavy Tessellation workloads.

 

[Wall Street]

Do you have any documentation for this ???

The GTX 1070 block diagram and specifications?

Same ROPs as the 1080 with one GPC disabled. The ROPs are the dark blue blocks next to the L2 cache in the chip diagram for the GTX 1080 and GTX 1070.