2016 nMP

[koyoot]

wrong calculcations.
1080 runs at stock 1.6, boost 1.73 and a lot of time it runs at higher 1.86 ghz.
(2100/1860 - 1) * 100% = 13 %, surprise.

Yes, but 13% faster than 1.6 GHz GTX 980 TI . Thats what I meant

155 FPS vs 136 FPS 18 FPS difference.

Hmmm, nope. Actual figures for that OC of GTX 980 TI water force is 1492 MHz.

So 2114 MHz 2560 CUDA core GPU is 13% faster than 1492 MHz 2816 CUDA core GPU. Yeah, IPC has decreased.

 

[netkas]

Ah right, misread that.

 

[Mago]

FP64 on GTX1080 & 1070 according wccftech is 1:16 FP32, so gtx1080 is 1/2 Tflop fp64

 

[koyoot]

FP64 on GTX1080 & 1070 according wccftech is 1:16 FP32, so gtx1080 is 1/tflop fp64

Not bad at all. Its more than Tahiti was able to put out, with its 1/4 ratio.

 

[Mago]

Interesting, I don't know if coincidence but GPGPU pricing still moreless constant at moreless 1000$ per FP64 TFLOP or a Buck per FP64 GFlop

With this in mind, and following my theory AMD Polaris XT shoul be at about 400 GFlop FP64
[doublepost=1463499314][/doublepost]@koyoot I edited my post GTX1080 FP64 performance is HALF Teraflop

 

[netkas]

Used titan and titan black has much better ratio then.

 

[koyoot]

Interesting, I don't know if coincidence but GPGPU pricing still moreless constant at moreless 1000$ per FP64 TFLOP or a Buck per FP64 GFlop

With this in mind, and following my theory AMD Polaris XT shoul be at about 400 GFlop FP64
[doublepost=1463499314][/doublepost]@koyoot I edited my post GTX1080 FP64 performance is HALF Teraflop

Yeah, my brain went on coffee break and I didn't see that there was something not right...

 

[Stacc]

FP64 on GTX1080 & 1070 according wccftech is 1:16 FP32, so gtx1080 is 1/2 Tflop fp64

Anandtech has it at 1:32.

Ars technica looked at the compute performance and its not as much of a slam dunk as the graphics performance. It beats the 980 Ti but not by as much as you would expect.

It sounds like the overclocking potential of the card is limited by the available power and not the thermals. I wonder what we will see when custom PCBs come to market with more than just an 8-pin connector.

Coming back around to the mac pro, we all expect to see Polaris 10 in it. There is room here for AMD to make a splash. If they can compete with the 1070 either on performance or price point (or both) then they could have a winner. Given that we also expect it to be more compute focused and has a lower power target it still makes sense to see this as the next mac pro card.

 

[netkas]

Wake me up when Amd's Vega released. The only interesting thing for second part of this year.

>It sounds like the overclocking potential of the card is limited by the available power and not the thermals. I wonder what we will see when custom PCBs come to market with more than just an 8-pin connector.

more like it's limited by power phases, 5+1(1 for memory) is not good for oc, something like 8 phases for gpu from custom cards would be better for oc

 

[AidenShaw]

Wake me up when Amd's Vega released. The only interesting thing for second part of this year.

What about the GTX 1080 ti and the Pascal-based Titan ?

 

[netkas]

What about the GTX 1080 ti and the Pascal-based Titan ?

it's early 2017 afaik

 

[AidenShaw]

it's early 2017 afaik

Nvidia said start of '17 for the GP100-based Tesla.

Maybe they're stockpiling GP104-600 chips for an October surprise.

 

[goMac]

What baffles me is this. The GPU has around 50% higher core clock than GTX 980 Ti, 50% more compute power than GTX 980 Ti, yet it is only 25% faster than GTX 980 Ti.

Yeah, this is what baffles me. Clock is much higher, large die shrink, and the gains are pretty.... meh. Ars had a pretty similar thought on their review. It's a nice jump otherwise, just... not the sort of jump you should be getting considering the spec changes.

 

[koyoot]

Actually, Titan will be based on GP102 - GP100 without FP64. However current rumors say that GP102 will not come until Q2 2017. The further in time 600mm2 die comes into consumer market without very high margin, the better for them. Like I have said, it is last node where building 600mm2 die sized GPU will be affordable. That is the reason why in the first place GP100 is released in HPC market.

Anyone want to pay 40000$ for Nvidia Titan class GPU? And that is reality on smaller nodes.

Yeah, this is what baffles me. Clock is much higher, large die shrink, and the gains are pretty.... meh. Ars had a pretty similar thought on their review. It's a nice jump otherwise, just... not the sort of jump you should be getting considering the spec changes.

I think I know why is it. GP100 has 64 cores per SM, GP104 has 128. So GP104 has less shared memory and register file per shader.
It looks like Nvidia did this to push clocks high enough without sacrificing efficiency.

Heh, And I was called AMD fanboy for saying that Pascal will not be as efficient as Maxwell.

All in all it looks like that despite having higher compute power, GTX 1070 will be slower than Titan X, and GTX 980 Ti.

 

[netkas]

Finally nvidia doesn't suck in async compute. It's not giving it much benefits but at least not taking down fps.

http://www.computerbase.de/2016-05/...ramm-ashes-of-the-singularity-async-compute_3

 

[Roykor]

Actually, Titan will be based on GP102 - GP100 without FP64. However current rumors say that GP102 will not come until Q2 2017. The further in time 600mm2 die comes into consumer market without very high margin, the better for them. Like I have said, it is last node where building 600mm2 die sized GPU will be affordable. That is the reason why in the first place GP100 is released in HPC market.

Anyone want to pay 40000$ for Nvidia Titan class GPU? And that is reality on smaller nodes.


I think I know why is it. GP100 has 64 cores per SM, GP104 has 128. So GP104 has less shared memory and register file per shader.
It looks like Nvidia did this to push clocks high enough without sacrificing efficiency.

Heh, And I was called AMD fanboy for saying that Pascal will not be as efficient as Maxwell.

All in all it looks like that despite having higher compute power, GTX 1070 will be slower than Titan X, and GTX 980 Ti.

Still, if they mange to get the 1070 some near the performance of the GTX980ti, and you pay 350 for the 1070, it will be a very interesting card to put in your workhorse.

 

[ManuelGomes]

A code to unlock 3 and 4 way SLI? WTF?
[doublepost=1463518006][/doublepost]Still no official word on DP ratio. I wonder why? I does indeed look to be 1:16 or 1:32, not sure.

 

[MacVidCards]

A code to unlock 3 and 4 way SLI? WTF?
[doublepost=1463518006][/doublepost]

Who cares, still runs rings around anything from AMD.

AMD never caught up with last gen high end, now they've been double-buried.

Best they've got is "sometime soon we'll have some midrange cards"

Sour grapes from the PR team?

Nvidia is so far ahead right now they can send their engineers on a month long nap with no worries.

AMD needs to get out of 2nd gear and get back in the game.

 

[Derpage]

Who cares, still runs rings around anything from AMD.

AMD never caught up with last gen high end, now they've been double-buried.

Best they've got is "sometime soon we'll have some midrange cards"

Sour grapes from the PR team?

Nvidia is so far ahead right now they can send their engineers on a month long nap with no worries.

AMD needs to get out of 2nd gear and get back in the game.

Absolutely nobody around here mentions that AMD left the high end market. The fanboys shifted the goal posts towards the perf/watt argument and have since argued that it's an economic choice and not a performance choice driving their decisions. They stick with this argument even after AMD openly admitted that they can not compete with Nvidia. Well, AMD realized they couldn't cut into Intel's market share either and are now back to attempting to make high end cards. I don't even know how to reply half the time to the nonsense spewed around here.

 

[koyoot]

A code to unlock 3 and 4 way SLI? WTF?
[doublepost=1463518006][/doublepost]Still no official word on DP ratio. I wonder why? I does indeed look to be 1:16 or 1:32, not sure.

Well, Nvidia has very generous audience. People will pay 100$ more for completely stock version of GPUs, so they can pay even more for a code to unlock 3 way or 4 way SLI.

 

[AidenShaw]

Well, Nvidia has very generous audience. People will pay 100$ more for completely stock version of GPUs, so they can pay even more for a code to unlock 3 way or 4 way SLI.

OMG - you made this argument on an Apple forum - that people will pay more for better stuff!

 

[koyoot]

OMG - you made this argument on an Apple forum - that people will pay more for better stuff!

You don't see that it is a CODE, and you could very well do not need that?

It is quite funny, now...

 

[linuxcooldude]

A code to unlock 3 and 4 way SLI? WTF?
[doublepost=1463518006][/doublepost]Still no official word on DP ratio. I wonder why? I does indeed look to be 1:16 or 1:32, not sure.

Must be that Nvidia tax.

 

[goMac]

Nvidia is so far ahead right now they can send their engineers on a month long nap with no worries.

Nvidia's performance per clock is so slow right now I don't think they should be napping.

This was a process change. Everyone was expecting faster. They could have slept through development of the 1080 and ended up with something twice as fast. That they didn't is a warning sign.

 

[koyoot]

Newly-added Polaris PCI IDs include 0x67C1, 0x67C2, 0x67C4, 0x67C7, 0x67C8, 0x67C9, 0x67CA, 0x67CC, and 0x67CF... Certainly more than before of the previous Polaris 10 PCI IDs being just 0x67C0 and 0x67DF

New Polaris 11 PCI IDs include 0x67E3, 0x67EF, and 0x67E7 to complement the current ones of 0x67E0, 0x67E1, 0x67E8, 0x67E9, 0x67EB, and 0x67FF.

Keep in mind the hardware vendors tend to reserve more PCI IDs than they necessarily end up releasing as products to market.

What the...?

Nvidia's performance per clock is so slow right now I don't think they should be napping.

This was a process change. Everyone was expecting faster. They could have slept through development of the 1080 and ended up with something twice as fast. That they didn't is a warning sign.

Compute performance on GP104 will suffer from smaller cache available for SM, and smaller registry file.

4K and high resolution VR and gaming experience also will suffer 6-10 months from now, due to narrow Memory bus. So far, the only truly worth attention GPU Nvidia released is GP100.

P.S. Haha, Nvidia provided 61 page book as a reviewers guide for GTX 1080 .