NVIDIA Pascal Thread

[ShintaiDK]

Lets assume realistically. 300mm2 die GPU with 4GPC accounting for 2560 CUDA cores total, with GDDR5X, 64 ROPs, and 1480 MHz boost clock.

We are looking at 200W GPU with 20-30% higher performance than GTX 980 Ti.

I think 200W is too much. Its going to be around 150W, max 175W. Its going to have ~50GB/sec faster memory than GTX980TI if GDDR5X.

 

[Glo.]

I think 200W is too much. Its going to be around 150W, max 175W.

Way too optimistic. The same type of hype of efficiency as is with people comparing Polaris 11 in performance to Fury X.

 

[Headfoot]

Sounds way to early to me to be declaring winners for Polaris 10 vs GP104

 

[ShintaiDK]

Way too optimistic. The same type of hype of efficiency as is with people comparing Polaris 11 in performance to Fury X.

A shrinked GTX980TI at 200W would beat the GTX980TI with 20-30% using a 256bit GDDR5X but instead of 384bit GDDR5.

 

[ShintaiDK]

Sounds way to early to me to be declaring winners for Polaris 10 vs GP104

Polaris 10 isn't near GP104 class in any way. Its closer to GP106.

The real competitor for GP104 is Vega 11.

 

[antihelten]

Way too optimistic. The same type of hype of efficiency as is with people comparing Polaris 11 in performance to Fury X.

No ones comparing Polaris 11 to Fury X, some people are comparing Polaris 10 to Fury X.

A shrinked GTX980TI at 200W would beat the GTX980TI with 20-30% using a 256bit GDDR5X but instead of 384bit GDDR5.

Are you saying that a reference 980 Ti is memory bottlenecked by 20-30%, or are you assuming that this shrunk 980 Ti would have a core clock 20-30% higher than reference (i.e. 1350-1450 MHz)?

 

[ShintaiDK]

Are you saying that a reference 980 Ti is memory bottlenecked by 20-30%, or are you assuming that this shrunk 980 Ti would have a core clock 20-30% higher than reference (i.e. 1350-1450 MHz)?

More bandwidth, higher clocks, less power.

 

[Timmah!]

Lets assume realistically. 300mm2 die GPU with 4GPC accounting for 2560 CUDA cores total, with GDDR5X, 64 ROPs, and 1480 MHz boost clock.

We are looking at 200W GPU with 20-30% higher performance than GTX 980 Ti.

7.5 TFLOPs of compute power. Hmmmm...

That for 1080.

1080Ti with 5GPC/3200 CC and 50 percent above 980Ti :-D

I know, wishful thinking.

 

[Glo.]

A shrinked GTX980TI at 200W would beat the GTX980TI with 20-30% using a 256bit GDDR5X but instead of 384bit GDDR5.

I am talking about power efficiency. We compare big cores, which are bigger than before - that is why Nvidia was able to fit only 30% more into GP100 at the same die size, as GM200, on a node that brings http://www.tsmc.com/english/dedicatedFoundry/technology/16nm.htm 2 times density. Also higher core clock is also accounting for higher TDP. 200W means: slightly lower, slightly more power draw. It can be either 195, for those that want it to be, or 205 for those who want it to be.

Overall, you did not disagreed with me anywhere about performance, apart from power efficiency. I do think we are looking at 20-30% faster GPU than GTX 980 Ti, but not in the same thermal envelope as GTX 980 was.

 

[Glo.]

Polaris 10 isn't near GP104 class in any way. Its closer to GP106.

The real competitor for GP104 is Vega 11.

You are wrong. 2560 GCN4 cores will be on par with 2560 Pascal Cores. That means, core-to clock. Already in DX12 we see similar thing with Maxwell GPUs, and GCN1.1, R9 390X competes with GTX 980 Ti. The difference is made by core clocks.

So Polaris 10 being 2560 GCN core GPU, clocked at 1150 MHz will be slightly slower than 2560 CUDA core, 1.48 GHz GPU from Nvidia.

The difference is what we were talking about: power efficiency. Polaris might be 125W GPU, that is replacing Pitcairn. And GP104 with 2560 CUDA cores can be either GTX 1080 or 1080 Ti, if the rumors are true.

 

[Kenmitch]

http://videocardz.com/59189/nvidia-geforce-gtx-10801070-cooler-shroud-taken-apart

1st impression....Fake.

 

[swilli89]

Polaris 10 isn't near GP104 class in any way. Its closer to GP106.

The real competitor for GP104 is Vega 11.

What do you base this assumption off of?

 

[antihelten]

More bandwidth, higher clocks, less power.

The reason I ask is because currently a factory overclocked 980 Ti can squezze out and extra 15% performance without any increase in bandwidth, and an extra 25% with only an 3% increase in memory bandwidth.

As such there isn't really any need for GDDR5X to reach a 20-30% improvement of a reference 980 Ti. All of this would then affect the TDP of course, and btw I agree with you that it should be possible to bring this down to 150-175W.

Overall, you did not disagreed with me anywhere about performance, apart from power efficiency. I do think we are looking at 20-30% faster GPU than GTX 980 Ti, but not in the same thermal envelope as GTX 980 was.

20-30% faster than a 980 Ti in the thermal envelope of a 980 would be an improvement in efficiency of roughly 60% which should be well within reach for a node jump. Remember a 680 was roughly 25-30% faster than a 580, whilst using ~25% less power, basically the same as what we're talking about here.

 

[MrTeal]

1st impression....Fake.

Yeah, until I see that at Computex I'm leaning that way. Besides it being garish, it really does look like someone at a small machine shop trolling people on the internet.

 

[ShintaiDK]

I am talking about power efficiency. We compare big cores, which are bigger than before - that is why Nvidia was able to fit only 30% more into GP100 at the same die size, as GM200, on a node that brings http://www.tsmc.com/english/dedicatedFoundry/technology/16nm.htm 2 times density. Also higher core clock is also accounting for higher TDP. 200W means: slightly lower, slightly more power draw. It can be either 195, for those that want it to be, or 205 for those who want it to be.

Overall, you did not disagreed with me anywhere about performance, apart from power efficiency. I do think we are looking at 20-30% faster GPU than GTX 980 Ti, but not in the same thermal envelope as GTX 980 was.

You are comparing apples and oranges with GP100 vs GM200. DP is one of the keys.

 

[ShintaiDK]

You are wrong. 2560 GCN4 cores will be on par with 2560 Pascal Cores. That means, core-to clock. Already in DX12 we see similar thing with Maxwell GPUs, and GCN1.1, R9 390X competes with GTX 980 Ti. The difference is made by core clocks.

So Polaris 10 being 2560 GCN core GPU, clocked at 1150 MHz will be slightly slower than 2560 CUDA core, 1.48 GHz GPU from Nvidia.

The difference is what we were talking about: power efficiency. Polaris might be 125W GPU, that is replacing Pitcairn. And GP104 with 2560 CUDA cores can be either GTX 1080 or 1080 Ti, if the rumors are true.

Hopefully you are not basing that on a handful of broken games. Besides wishful dreams of GCN 1.3 performance. There was no limit how much better GCN 1.2 was over GCN 1.1 either...until release.

Also its funny how your Polaris chip had no problem reaching 125W, while 200W was needed for GP104

 

[Glo.]

20-30% faster than a 980 Ti in the thermal envelope of a 980 would be an improvement in efficiency of roughly 60% which should be well within reach for a node jump. Remember a 680 was roughly 25-30% faster than a 580, whilst using ~25% less power, basically the same as what we're talking about here.

Erm... that is Exactly what I said .

If 2560 CUDA core 1.48 GHz GPU will use around 200W of power, and will be 20-30% faster than GTX 980 Ti, it will mean that it will use 20% less power, than predecessor .

 

[Sweepr]

If there was any hope for Vega launching in late 2016, not anymore based on AMD's latest roadmap:

Earnings Call tomorrow.

 

[antihelten]

Erm... that is Exactly what I said .

If 2560 CUDA core 1.48 GHz GPU will use around 200W of power, and will be 20-30% faster than GTX 980 Ti, it will mean that it will use 20% less power, than predecessor .

No it doesn't, a 980 Ti does not use 250W, that's it's TDP not it's power usage.

211W for a reference 980 Ti
156W for a reference 980, 26% less.

 

[USER8000]

If there was any hope for Vega launching in late 2016, not anymore based on AMD's latest roadmap:

It is from the same month as the AMD Capsaicin event it appears.

 

[xorbe]

Yeah, until I see that at Computex I'm leaning that way. Besides it being garish, it really does look like someone at a small machine shop trolling people on the internet.

Yeah, strange leak. I think that's the CNC guy saying, hint hint.

 

[antihelten]

Yeah, strangle leak. I think that's the CNC guy saying, hint hint.

Based on the pics it would look like it's injection molded, not made on a CNC mill (there's a mold in one of the pics).

Faking an injection molded cooler shroud should be well outside the capability of most people.

 

[mohit9206]

How long before we get a graphics card from AMD or Nvdia as powerful as a Gtx960 for under $100? The entry level video card market has been stagnant for far too long and Intel lighting fire under AMD's and Nvdia's arses hasn't led to anything.

 

[MrTeal]

Based on the pics it would look like it's injection molded, not made on a CNC mill (there's a mold in one of the pics).

Faking an injection molded cooler shroud should be well outside the capability of most people.

Depends on your capabilities, I guess. We have a single machinist to support our small (half-dozen Engineer) business unit, and we have an EDM machine for mold making. They wouldn't be common in a hobbyist garage, but they're far from uncommon. If I were so motivated I could have a shroud made up with "490X" written on it and shaped like a giant penis, though the company might not find that to be a particularly wise use of resources.

That being said, it would be a dedicated troll. Props to him if it is.

 

[airfathaaaaa]

If there was any hope for Vega launching in late 2016, not anymore based on AMD's latest roadmap:

Earnings Call tomorrow.

there was a hope to begin with? they clearly stated vega would be an early 20178 card