NVIDIA Kepler speculation thread

[AlphaWolf]

Look to the forum options in your control panel! You can set to which size pictures should be rescaled (default is a width of 800 pixels and unrestricted height).

Spoiler

/OT

Sadly that doesn't stop the images from loading. This thread has become decidedly painful on mobile unless I disable images or put certain users on ignore.

 

[Wynix]

http://www.anandtech.com/show/6838/nvidia-geforce-gtx-650-ti-boost-review-/14



Now considering you can pick up a 7850 1GB for the same performance and less cash, and the Reloaded bundle...

If you throw overclocking into the equation then the AMD card looks even better, the 7850 is clocked super low ~860mhz core, every 7850 can hit 1050 no sweat.

 

[Blazkowicz]

Sadly that doesn't stop the images from loading. This thread has become decidedly painful on mobile unless I disable images or put certain users on ignore.

I remember that on the Opera browser (for desktops, old version) you could trivially access a button or menu option to disable/enable images.
Dunno if Opera mobile would help the same way but anyway you need some software solution to not load images when you come on this thread or on forum.beyond3D.com

 

[DSC]

http://www.geforce.com/whats-new/ar...80-powered-unreal-engine-4-tech-demo-unveiled

http://www.youtube.com/watch?v=dO2rM-l-vdQ

UE4 Infiltrator real-time demo running on a single GTX 680.

 

[A1xLLcqAgt0qc2RyMz0y]

GeForce GTX Titan, GeForce GTX 690, Radeon HD 7990 (HD 7970 CrossFire)

Frame Rating: GeForce GTX Titan, GeForce GTX 690, Radeon HD 7990 (HD 7970 CrossFire)

http://www.pcper.com/reviews/Graphi...eForce-GTX-690-Radeon-HD-7990-HD-7970-CrossFi

The Titan looks even better than the GTX 690 than some of the earlier reviews showed.

Benefits of Single GPUs

The GeForce GTX 690 and the GeForce GTX Titan paint an interesting picture if you just consider them against each other. Both cards will cost you a cool $999, but offer very different hardware configurations. The GTX 690 is a dual-GK104 product with 4GB of total frame buffer and 2GB for each GPU; essentially slightly slower GTX 680s in SLI. The GTX Titan is a totally different beast with a single, larger GK110 GPU (2,688 cores in a single GPU, 1,536 in each GK104) and a larger 6GB frame buffer. When it launched many people complained that the GTX Titan was overpriced because it offered lower performance than the GTX 690 - which shows to be the case in most of our FRAPS benchmarks and even the observed FPS data.

What you might not have seen before though is what happens when we crank up the resolution to triple monitor setups - our 5760x1080 benchmarks. In those cases, pretty much across the board, we found that the GTX Titan was able to produce much more smooth and consistent frame times compared to the GTX 690. This is likely a combination of both the large frame buffer of the Titan, (triple the memory) as well as not having the need to worry about balancing frames across multiple GPUs and doing any kind of frame metering. When the GTX 690 is running games at higher settings on 5760x1080, it struggles to keep up.

NVIDIA's GeForce GTX 690 and GTX Titan present another debate though on the issue of single versus multi-GPU options. In our testing the GTX 690 definitely introduces more frame time variance than the GTX Titan, but the cards are so fast as it is that running the GTX 690 at single monitor resolutions like 1920x1080 and 2560x1440 don't cause a big enough problem to be a factor, thus giving it the edge because of the higher average frame rates. The GTX Titan is a powerful card and performs admirably in the single display testing but it really stands away from the GTX 690 on multi-monitor resolutions like 5760x1080 where the 6GB frame buffer can help a TON. Not having to worry about moving frames between GPUs at that resolution also helps produce a smoother animation as well. If you are buying a $1000 card and you think you might want to run NVIDIA Surround, then the GTX Titan is your better solution.

 

[Wynix]

Frame Rating: GeForce GTX Titan, GeForce GTX 690, Radeon HD 7990 (HD 7970 CrossFire)

http://www.pcper.com/reviews/Graphi...eForce-GTX-690-Radeon-HD-7990-HD-7970-CrossFi

The Titan looks even better than the GTX 690 than some of the earlier reviews showed.

Not really no; this just shows the already known flaws of running multiple cards versus a single card, nothing new here at all.

The thing about 6GB being helpful for multi monitors is baseless, nothing in the review suggested that having 6GB improved performance, the review on the 6GB 7970 came to that conclusion(toxic-hd-7970-eyefinity-6gb), so i don't know how the reviewer can just throw that in there without any proof.

Everything about the Titan having lower latency on multiple monitors can be attributed to it being a single GPU card compared to SLI and nothing to do with the amount of memory it has.

Ignore this part as i skipped the first page.
Also they are using fraps, Anand have pointed out that it's a very imperfect way of recording frame latency(Fraps).

 

[silent_guy]

Also they are using fraps, Anand have pointed out that it's a very imperfect way of recording frame latency(Fraps).

Well, at least this comment settles the question about whether or not you read the linked pcper article.

 

[Wynix]

Well, at least this comment settles the question about whether or not you read the linked pcper article.

Not at all.

 

[DSC]

http://nvidianews.nvidia.com/Releas...mance-Out-of-Notebooks-Automatically-96c.aspx

http://www.geforce.com/whats-new/articles/introducing-the-new-geforce-700m-series-gpus

Nvidia has announced the new GeForce 700M GPUs.

GeForce GT 750M, GeForce GT 745M, and GeForce GT 740M GPUs for the performance segment,
GeForce GT 735M and GeForce GT 720M GPUs for the mainstream segment.

I hope this is all using the new GK208 GPU with Compute Capability 3.5.

http://www.geforce.com/hardware/notebook-gpus/geforce-gt-750m

http://www.geforce.com/hardware/notebook-gpus/geforce-gt-745m

http://www.geforce.com/hardware/notebook-gpus/geforce-gt-740m

http://www.geforce.com/hardware/notebook-gpus/geforce-gt-735m

http://www.geforce.com/hardware/notebook-gpus/geforce-gt-720m

 

[silent_guy]

Not at all.

?
The whole point of that article is to use a frame capturer and compare it against FRAPS to show how the latter is insufficient.

 

[DSC]

http://www.anandtech.com/show/6873/nvidias-geforce-700m-family-full-details-and-specs

Looks like only the 735M and 730M is GK208 GPU.

 

[AnarchX]

Package shots:
http://www.geforce.com/Active/en_US...gt-745m/geforce-gt-745m-F_gallery_preview.png GK107?
http://www.geforce.com/Active/en_US...gt-735m/geforce-gt-735m-F_gallery_preview.png GK208?

Comparing them based on die markings (circles, triangle, squares), GK208 seems only be around 70mm².

 

[tviceman]

Package shots:
http://www.geforce.com/Active/en_US...gt-745m/geforce-gt-745m-F_gallery_preview.png GK107?
http://www.geforce.com/Active/en_US...gt-735m/geforce-gt-735m-F_gallery_preview.png GK208?

Comparing them based on die markings (circles, triangle, squares), GK208 seems only be around 70mm².

As far as my understanding, GK208 is 2 SMX's with 1 64 bit memory controller. So removing the memory interface and controller, and perhaps (maybe) slightly tweaking the SMX design to save space, shrinking the die size from GK107's 118 mm^2. If it is only around 70mm^2, that is quite a large reduction. In the neighborhood of 40%.

 

[tviceman]

http://nvidianews.nvidia.com/Releas...mance-Out-of-Notebooks-Automatically-96c.aspx

http://www.geforce.com/whats-new/articles/introducing-the-new-geforce-700m-series-gpus

Nvidia has announced the new GeForce 700M GPUs.

GeForce GT 750M, GeForce GT 745M, and GeForce GT 740M GPUs for the performance segment,
GeForce GT 735M and GeForce GT 720M GPUs for the mainstream segment.

I hope this is all using the new GK208 GPU with Compute Capability 3.5.

http://www.geforce.com/hardware/notebook-gpus/geforce-gt-750m

http://www.geforce.com/hardware/notebook-gpus/geforce-gt-745m

http://www.geforce.com/hardware/notebook-gpus/geforce-gt-740m

http://www.geforce.com/hardware/notebook-gpus/geforce-gt-735m

http://www.geforce.com/hardware/notebook-gpus/geforce-gt-720m

The 750m has a nice bump in performance over the 660m - 25% more bandwidth and about 16% more fill rate. Should be a smaller power envelope as well. Even if these aren't tweaked chips, node process maturity helped quite a bit.

 

[iMacmatician]

As far as my understanding, GK208 is 2 SMX's with 1 64 bit memory controller. So removing the memory interface and controller, and perhaps (maybe) slightly tweaking the SMX design to save space, shrinking the die size from GK107's 118 mm^2. If it is only around 70mm^2, that is quite a large reduction. In the neighborhood of 40%.

Well now that you mention it, that is a considerable decrease in die area (GF108 to GF117 had a slightly greater die size reduction but that was from a process shrink and it kept the 128-bit bus). If GK208 has SMX optimizations over GK107, I wonder how much space can be saved from GK104 and GK106 if they undergo the same treatment.

 

[mczak]

Package shots:
http://www.geforce.com/Active/en_US...gt-745m/geforce-gt-745m-F_gallery_preview.png GK107?
http://www.geforce.com/Active/en_US...gt-735m/geforce-gt-735m-F_gallery_preview.png GK208?

Comparing them based on die markings (circles, triangle, squares), GK208 seems only be around 70mm².

Doesn't quite match my pixel comparisons. Based on the bga points the scale factor seems to be very close to 1.25. And going from that (gk208 shot also seems to include more border around the logo) gk107 seems to be about 10% larger in x direction, 15% larger in y direction, which would make it below 30% larger. Ok accounting for inaccuracies it could well be a bit more, but that would make gk208 about 90mm².
I think that would be more in line what's expected for dropping a MC/ROP (together with gddr5 support presumably, and maybe no display outputs neither?). Still quite a large reduction actually, though almost certainly not worth it considering perf/w and perf/area (and it will probably get blown away by the similar sized (85mm²) 128bit mars/oland in practice except extremely shader limited scenarios such as 3dmark11).

I don't quite get nvidia's numbers for comparison against HD4000 anyway. That factor 3 on the low end seems to be very far from reality, should be more like 1.5. I guess they took the lowest clocked HD4000 they could find anywhere, outfitted it with ddr3-800 single channel and called it a day. Or maybe the benchmarks (and settings) chosen really are that indifferent to memory bandwidth. That the numbers for a 64bit ddr3 and a (presumably gddr5 even) 128bit version are so close should already tell you how much those numbers are worth.

 

[AnarchX]

You are right, there was a calculation error.

 

[tviceman]

Doesn't quite match my pixel comparisons. Based on the bga points the scale factor seems to be very close to 1.25. And going from that (gk208 shot also seems to include more border around the logo) gk107 seems to be about 10% larger in x direction, 15% larger in y direction, which would make it below 30% larger. Ok accounting for inaccuracies it could well be a bit more, but that would make gk208 about 90mm².

That sounds way more realistic. Still, at 90mm^2, that is nearly 30mm^2 shaved off GK107 by removing 1 memory controller and memory interface. I still think there had to be some space saving measures taken elsewhere to shave off 30mm^2. Perhaps the SMX modules on GK208 are essentially what will be on Tegra 5 (but likely on 20nm). I know Tegra 5's GPU clock won't be as high as anything discrete cards have, but still it will be a quite a large step up in GPU power.



ANYWAYS, TSMC is ahead of schedule with 20nm: http://focustaiwan.tw/news/atod/201304010042.aspx That's good news.

 

[UniversalTruth]

ANYWAYS, TSMC is ahead of schedule with 20nm: http://focustaiwan.tw/news/atod/201304010042.aspx That's good news.

Hmm... and what about this

TSMC and ARM Tape-Out First ARM Cortex-A57 Processor on 16 nm FinFET Technology

 

[mczak]

That sounds way more realistic. Still, at 90mm^2, that is nearly 30mm^2 shaved off GK107 by removing 1 memory controller and memory interface. I still think there had to be some space saving measures taken elsewhere to shave off 30mm^2. Perhaps the SMX modules on GK208 are essentially what will be on Tegra 5 (but likely on 20nm).

It is quite possible there's some small changes somewhere to shave off some area. But if it doesn't support gddr5 at all then even the remaining 64bit gddr5 channel should save quite some area on the PHY (if you compare bonaire with cape verde the former seems to have an astonishing large increase on the PHYs and that was only to support 6Ghz (or possibly 7Ghz) instead of 4.5Ghz gddr5). Also, I haven't seen if there's any display outputs at all - if they aren't present that will save some area as well. Without a die shot it seems quite impossible to tell if the area savings would be due to optimizations or indeed mostly just by cutting stuff.