NVIDIA Pascal Thread

[Leadbox]

Cant wait for the performance/watt benefits.

But for raw performance people may end up disappointed.

They're putting back in the compute stuff they neutered for maxwell's perf/watt advantage because they need quadros and teslas out of pascal, the perf/watt might just mirror the raw perf that you worry about.

 

[ShintaiDK]

They're putting back in the compute stuff they neutered for maxwell's perf/watt advantage because they need quadros and teslas out of pascal, the perf/watt might just mirror the raw perf that you worry about.

The lack of raw performance comes from die sizes.

Take AMD that we know more about with 120 and 230? mm2.

 

[Silverforce11]

If these small chips can match top-end stuff on performance, that's pretty good.

 

[Mahigan]

Great post Aracnothronic, really exposes the double standards here.



No, according to the leaks different Pascal products will be demonstrated in April @ GTC and May, with possible launch @ May-July timeframe. Completely different from the FUD we've heard from the usual suspects: 'AMD/Polaris has a 6 month advantage to market'.

Yeah the mobile parts first followed by the 1070 and 1080. The 1080 Ti and Titan cards won't be arriving for some time though.

It's the same for AMD. Polaris 10/11 first with Vega 10/11 later.

 

[Sweepr]

Yeah the mobile parts first followed by the 1070 and 1080. The 1080 Ti and Titan cards won't be arriving for some time though.

It's the same for AMD. Polaris 10/11 first with Vega 10/11 later.

That's not what the latest Bench-life rumours indicates, but it wouldn't surprise me if mobile GPUs were first. I think they're going to unveil both GP104 and GP106 around the same time.

 

[JDG1980]

No, according to the leaks different Pascal products will be demonstrated in April @ GTC and May, with possible launch @ May-July timeframe. Completely different from the FUD we've heard from the usual suspects: 'AMD/Polaris has a 6 month advantage to market'.

The difference is that from Nvidia, all we have is rumors and alleged leaks. Nothing specific, nothing concrete. We haven't seen an actual Pascal chip even shown in public, much less demonstrated in a working system. I think we can safely say that Nvidia didn't have viable GP106 silicon in early January, or else they would have used that at the Drive PX 2 demo instead of the GM204 modules.

AMD demonstrated a working Polaris chip months ago, and we've also seen specific confirmation from official company representatives that Polaris will be released mid-year (which I assume could be any time from May to August).

Sure, it's possible that Nvidia has just been playing its cards close to the vest and is actually much closer to release than anyone thinks. But Charlie Demerjian (who blew the whistle on Bumpgate and Woodscrewgate) is fairly sure Pascal will be late, so I trust his judgment on this matter.

 

[Sweepr]

Sure, it's possible that Nvidia has just been playing its cards close to the vest and is actually much closer to release than anyone thinks. But Charlie Demerjian (who blew the whistle on Bumpgate and Woodscrewgate) is fairly sure Pascal will be late, so I trust his judgment on this matter.

I wouldn't trust any info from him regarding NVIDIA's plans/products, we all know his feelings about the company.

 

[Timmah!]

Would you guy be interested in something like this?

Geforce 'GTX 1080'
- +$500
- Full GP104 die
- Relatively small die and significant OC headroom (>20%)
- 15-25% faster than Geforce GTX980 Ti
- 256-bit / 512-bit GDDR5, new model @ H2 with GDDR5X

Geforce 'GTX 1070'
- +$400
- Relatively small die and significant OC headroom (>20%)
- Slightly faster than Geforce GTX980 Ti
- 256-bit / 512-bit GDDR5, new model @ H2 with GDDR5X

Thats not good enough to me. I could have bought 980Ti on Christmas...in the end i decided against it cause of Pascal, but if i hypothetically did, why would i want to drop another +500 EUROs (after already spending cca 750 on 980Ti) on just slightly higher performance about half year later?

 

[raghu78]

Cant wait for the performance/watt benefits.

But for raw performance people may end up disappointed.

I don't think so. Both Polaris and Pascal are going to be much better DX12 architectures than their predecessors. I also think FINFET is going to bring higher core frequencies for AMD who really struggled towards the end at 28nm as they pushed transistor density to the limit with Fiji thus severely limiting OC. Nvidia kicked serious ass with Maxwell clocks and OC and will atleast maintain or do better with Pascal.

If we look at past process node + architecture transitions the GTX 660 got close to GTX 580 and GTX 680 beat GTX 580 by 30-35%. I think something similar or better can be expected.

My guess is GP104 with 4096cc, 256 bit GDDR5X, 8 GB, 128 ROPs will come in at around 350 sq mm and kick some serious ass against 980 Ti. btw I am predicting a Maxwell type launch cadence with tiny 100-120 sq mm GPUs in July and the 300+ sq mm GPU in late Q3 or early Q4.

 

[Timmah!]

I don't think so. Both Polaris and Pascal are going to be much better DX12 architectures than their predecessors. I also think FINFET is going to bring higher core frequencies for AMD who really struggled towards the end at 28nm as they pushed transistor density to the limit with Fiji thus severely limiting OC. Nvidia kicked serious ass with Maxwell clocks and OC and will atleast maintain or do better with Pascal.

If we look at past process node + architecture transitions the GTX 660 got close to GTX 580 and GTX 680 beat GTX 580 by 30-35%. I think something similar or better can be expected.

My guess is GP104 with 4096cc, 256 bit GDDR5X, 8 GB, 128 ROPs will come in at around 350 sq mm and kick some serious ass against 980 Ti. btw I am predicting a Maxwell type launch cadence with tiny 100-120 sq mm GPUs in July and the 300+ sq mm GPU in late Q3 or early Q4.

I doubt GP104 is going to have 4096 cores. Thats too good to be true.

 

[ShintaiDK]

I doubt GP104 is going to have 4096 cores. Thats too good to be true.

Agree, I dont believe that either.

 

[raghu78]

I doubt GP104 is going to have 4096 cores. Thats too good to be true.

I don't think so. Nvidia need a good perf jump to sell the new flagship and ask enthusiast consumers to cough up USD 600 - 650. GTX 580 was a 520 sq mm die with FP64 focus while GTX 680 was 294 sq mm. GTX Titan X is a 601 sq mm die with no fp64 focus. So I think Nvidia are going to have to get up to 350 sq mm die to make the next gen flagship a good bump up from Titan X. GM204 was 398 sq mm with a portion of that included for fixed function like PCI-E controller, video controller and codecs. Doubling the cuda cores and associated logic alone while maintaining a 256 bit GDDR5X controller should bring that die size down to 350 sq mm. Its definitely possible. With a little bit of architectural improvements added in the GP104 should be impressive.

 

[Timmah!]

I don't think so. Nvidia need a good perf jump to sell the new flagship and ask enthusiast consumers to cough up USD 600 - 650. GTX 580 was a 520 sq mm die with FP64 focus while GTX 680 was 294 sq mm. GTX Titan X is a 601 sq mm die with no fp64 focus. So I think Nvidia are going to have to get up to 350 sq mm die to make the next gen flagship a good bump up from Titan X. GM204 was 398 sq mm with a portion of that included for fixed function like PCI-E controller, video controller and codecs. Doubling the cuda cores and associated logic alone while maintaining a 256 bit GDDR5X controller should bring that die size down to 350 sq mm. Its definitely possible. With a little bit of architectural improvements added in the GP104 should be impressive.

Dont get me wrong, i wish it turned out to be truth, but given the fact gtx980 had less cores than previous flagship (780Ti) and still sold like crazy, even to people who actually owned 780Ti, i am not holding my breath. If GP104 narrowly beats 980Ti with just 2500 shaders, you can count on it, thats the amount it will have...and not 4000...that would happen only if there were no other architectural improvements, so the only way to beat big Maxwell would be to have more cores.

At least thats my opinion. If GP104 really had 4096 cores, i might actually reconsider getting it. But that would mean 133 percent of Titan X core count, add those IPC improvements on top of it and perhaps higher clocks...and we may look at 40-50 percent more performance than Titan X/980Ti....do you think that can happen?

 

[Sweepr]

Thats not good enough to me. I could have bought 980Ti on Christmas...in the end i decided against it cause of Pascal, but if i hypothetically did, why would i want to drop another +500 EUROs (after already spending cca 750 on 980Ti) on just slightly higher performance about half year later?

Personally what I'm expecting from GP104 is a miniaturized version of GM200, with 3072 cores. Coupled with some architectural improvements, core clocks north of 1 GHz, 8 GB GDDR5/GDDR5X, <150W TDP and a competitive price relative to Polaris it could become a very interesting product. Many OCed Geforce GTX 980 Ti users probably wouldn't upgrade, but then again I bet most of them are looking forward to big Pascal anyway.

 

[raghu78]

Dont get me wrong, i wish it turned out to be truth, but given the fact gtx980 had less cores than previous flagship (780Ti) and still sold like crazy, even to people who actually owned 780Ti, i am not holding my breath. If GP104 narrowly beats 980Ti with just 2500 shaders, you can count on it, thats the amount it will have...and not 4000...that would happen only if there were no other architectural improvements, so the only way to beat big Maxwell would be to have more cores.

At least thats my opinion. If GP104 really had 4096 cores, i might actually reconsider getting it. But that would mean 133 percent of Titan X core count, add those IPC improvements on top of it and perhaps higher clocks...and we may look at 40-50 percent more performance than Titan X/980Ti....do you think that can happen?

GM204 aka GTX 980 was on the same 28nm node as GK110 aka GTX 780 Ti. GP104 comes with a full process node shrink (in fact slightly better in terms of power and efficiency gains).

http://www.tsmc.com/english/dedicatedFoundry/technology/16nm.htm

"TSMC's 16FF+ (FinFET Plus) technology can provide above 65 percent higher speed, around 2 times the density, or 70 percent less power than its 28HPM technology. "

I can confidently say 28HPM to 16FF+ is the biggest leap in transistor performance and power efficiency compared to any previous TSMC node transition.

This massive process node jump combined with architectural improvements means I expect 40% perf over Titan X as very reasonable. The lead should be even better in DX12 games.

 

[ShintaiDK]

Personally what I'm expecting from GP104 is a miniaturized version of GM200, with 3072 cores. Coupled with some architectural improvements, core clocks north of 1 GHz, 8 GB GDDR5/GDDR5X, <150W TDP and a competitive price relative to Polaris it could become a very interesting product. Many OCed Geforce GTX 980 Ti users probably wouldn't upgrade, but then again I bet most of them are looking forward to big Pascal anyway.

Likewise

 

[nvgpu]

Nvidia will promote GP104 upgrade to GM204 users, just like GM204 was promoted to GK104.

http://international.download.nvidi...rce-gtx-980-2560x1600-high-aa-performance.png

If you already have the flagship GM200 GPU, you wouldn't upgrade to GP104 anyway since it's not as big an upgrade, same with GK110 users wouldn't upgrade to GM204 anyway.

I expect TITAN X & GTX 980 Ti will be discontinued immediately when GP104 launches just like GTX 780 Ti was discontinued after GTX 980 was released.

Due to the launch of the GTX 980 and GTX 970, NVIDIA&#8217;s product lineup will be changing to accommodate these cards. GTX 780 Ti, GTX 780, and GTX 770 are all being discontinued; their replacements offer better performance at better prices for lower power consumption.

 

[tential]

Nvidia will promote GP104 upgrade to GM204 users, just like GM204 was promoted to GK104.

http://international.download.nvidi...rce-gtx-980-2560x1600-high-aa-performance.png

If you already have the flagship GM200 GPU, you wouldn't upgrade to GP104 anyway since it's not as big an upgrade, same with GK110 users wouldn't upgrade to GM204 anyway

Not sure if you're serious or just kidding...
Poor GK110...
The architecture that just couldn't.

 

[jpiniero]

Nvidia will promote GP104 upgrade to GM204 users, just like GM204 was promoted to GK104.
.

Think of it this way:

- The 1070 would be $399, slightly slower than the 980 Ti but still 30% faster than the 970. So yeah the target would be people who have the 970 and Kepler 1 and anything earlier. You'd get optimized DX12 and 8 GB of memory instead of 3.5 GB. You know nVidia will gimp it somehow but I guess we will have to see.

- The 1080 would be $649 and 20-25% faster than the 980 Ti. The target would be people with the 980 and Kepler 2... and of course people with GM200 but now their e-penis is flaccid.

And of course mobile versions would be a big improvement.

 

[nvgpu]

That's way off the mark, GTX 680 was $499 and GTX 980 was $549.

Full GP104 will be $499 - $549.

$649 - $699 pretty much reserved for flagship GP100.

 

[ShintaiDK]

I expect 599 for a full GP104. And I expect all 14/16nm parts to up the prices. A 799$ NVidia/AMD 980TI/Fury X replacement is what I expect as well.

 

[dark zero]

That's way off the mark, GTX 680 was $499 and GTX 980 was $549.

Full GP104 will be $499 - $549.

$649 - $699 pretty much reserved for flagship GP100.

Sadly, new materials and new process will increase the cost.
Also inflation is there.

So:

Full GP 104 must be around $549 - $599 tier
While the Flagship can be over $700 range.

Similar things apply to Polaris.

The interesting thing will be at which cost will be the GTX 950 sucessor. Definately will be at 16 nm and will have pretty great features.

 

[nvgpu]

GP107 will deliver the features that GM107 didn't have, fixed function HEVC Main/Main10 hardware decoder, VP9 hardware decoder & HDMI 2.0/2.0a support for the HTPC market.

The GTX 950 75W no PCIe power connector cards from Asus can fill that market for now though.

 

[linkgoron]

I think that talks about prices are premature.

If nvidia hard-launches first or Polaris is un-competitive then prices will be high/very high. Only competition creates low prices.

I recall this:
http://www.cnet.com/news/nvidia-cuts-prices-on-gtx-260-280-graphics-boards/
http://www.anandtech.com/show/7465/...-gtx-780770-price-cuts-gtx-780-ti-launch-date

We could see a repeat, if indeed nvidia launches first with some margin.

 

[crisium]

I expect 599 for a full GP104. And I expect all 14/16nm parts to up the prices. A 799$ NVidia/AMD 980TI/Fury X replacement is what I expect as well.

I see those prices most likely as well, though +/- $50 is possible.