November 14th, 2012

New NVIDIA Roadmap Reveals Maxwell Coming in 2014, GeForce GTX 780 Details

GeForce GTX 780

Over the past two days rumors regarding the upcoming graphics cards from NVIDIA have surfaced. First of all, we have a new roadmap, which doesn’t really reveal that much new information, but it’s definitely updated. It is also worth bringing up the specs of GeForce GTX 780 revealed by German website Chip.de.

Previous NVIDIA roadmap mentioned Kepler architecture to be released on 2011, but we all know they were not launched 2 years ago, but this year, so it was quite out-dated. The new one does not mention any Kepler Refresh GPUs for 2013, but nor does it mention a Fermi Refresh.

So it seems that the completely redesigned silicon, also known as Maxwell, will be launched in 2014. Redesign means a new approach in dealing with the computing power of the graphics processor, which would now be supported by an additional ARM core (which is probably a ARMv8 64-bit processor).

Maxwell GPUs would offer a significant performance boost in computing power per Watt — 2.6x times faster than the current Kepler generation per Watt. But before you get too excited about this fact, it’s worth saying that it does not mean that the graphics cards will be 2.6x faster overall. It means that the GPUs will be smaller (20nm), will generate less heat and consume less power.

GeForce GTX 700 Series – Kepler Refresh

But before we learn any details about Maxwell GPUs, there’s still a Kepler Refresh to come. These GPUs will likely launch in the first quarter of 2013. The most anticipated one is obviously a derivative of GK110 silicon. However, NVIDIA will not launch the same GPU known from Tesla as a GeForce card. Some say it’s just a castrated version of it, some say it’s a completely new GPU.

NVIDIA GeForce GTX 780 Specification

GPU: GK2xx
Base Clock: 1100 MHz
Boost Clock: 1150 MHz
Memory Clock: 1625 MHz (6500 MHz effective)
Memory type: 3GB GDDR5
Memory Interface: 384-bit

According to Chip.de, the GeForce GTX 780, which would supposedly use this processor, may feature 2304 CUDA cores. The site also reveals that the card would use 3GB GDDR5 memory across a 384-bit interface. Nevertheless, our sources are less enthusiastic about this card. From what we know, it would have 1920 CUDA cores (2110 tops), 160 TMUs and 48 ROPs. Needless to say, probably the most certain information right now is the 384-bit interface. The site also mentions the exact clock speeds, but it seems that these are pure speculation, since NVIDIA does not usually launch their cards with such round clock numbers. Anyway, from what has been said so far, the GPU would be clocked at 1100/1150 MHz (for base and boost respectively), and the effective memory clock is 6500 MHz.

These specs suggest that the memory bandwidth would increase from 192 GB/s to 312 GB/s (that’s 62% increase). The texture fill-rate would be around 170.6 GT/s (32% increase) and pixel fill-rate 52.8 GP/s (64% increase)

The Kepler Refresh Naming

It’s still unclear whether NVIDIA will launch their new GPUs with the GK1xx or GK2xx naming. However our sources and the leak from previous week suggest that the Kepler Refresh is using a GK2xx scheme. What that means is there is no GK110 or GK114 powered GeForce graphics card coming.

GeForce GTX 780 Launch Date

The GTX 780 is currently planned for February/March 2013. First engineering samples should be available in January. The release date of GTX 780 might be changed depending of the sales of current Kepler cards. However the information that GeForce GTX 780 might arrive this year is a complete nonsense.

Source: WCCFTECH

Author: WhyCry

ZoSo
MOAR CUDAZ!!!!
Simonsays
Will they run into manufacturing problems like they did with GK104?
ZoSo
“Interestingly NVIDIA tells us that their yields are terrific – a statement backed up in their latest financial statement – so the problem NVIDIA is facing appears to be demand and allocation rather than manufacturing.” – AnAndTech
http://www.anandtech.com/show/6446/nvidia-launches-tesla-k20-k20x-gk110-arrives-at-last/2
http://videocardz.com/ WhyCry (VideoCardz.com)
It’s still 28nm, the problems were caused by the transition into smaller fabrication process.
Simonsays
Really Nvidia had problems bringing Kepler to the Market that suggests manufacturing problems and not demand/yield issues. Nvidia blamed TSMC, yet all the other big game players AMD, Qualcomm never has any issues.
http://semiaccurate.com/2012/05/01/why-cant-nvidia-supply-keplergk104gtx680/
ZoSo
My linked article is talking about actual GK110 yields right now, not 6 months ago like that SI article. I don’t know how much more of a direct answer you could get.
BestJinjo
Your math is all over the place ;)
1) Memory bandwidth = 192 GB/s to 312 GB/s (that’s 38% increase). —> 312 / 192 would be a 62.5% increase
2) Texture fill-rate would be around 170.6 GT/s (24.5% increase) —> GTX680 had 128.8 –> would be 170.6 / 128.8 or a 32.4% increase
3) Pixel fill-rate 52.8 GP/s (39% increase) —> GTX680 had 32.2 —> 52.8 / 32.2 or a 64% increase
I think you should also note that those are Speculation Specifications. The Chip.de article states that it’s just their estimate (i.e., they are making it up out of thin air without sources).
The roadmap for Kepler 2012 and Maxwell for 2014 was known a long time ago, at least since July 6th, 2011 and it was already known then Maxwell was a 20 or 22nm part, not 28nm, hence the delay from 2013 to 2014:
http://news.softpedia.com/news/Nvidia-Kepler-Will-Arrive-in-2012-Maxwell-in-2014-Says-Company-Slide-210023.shtml
BestJinjo
He is right though about GTX600 series launch. While yields are fine now, NV had to scrap GK100/110 for consumer GPUs this year and they started shipping low volumes of GK110 in late October to fill the spring order backlog for HPC customers. Now they are selling every K10/K20 they can make. They are wafer constrained at this point.
http://videocardz.com/ WhyCry (VideoCardz.com)
Thanks for noticing my errors.
I made this post to provoke some discussion about the specs, without assuming that any of those which float around the web are actually true.
You are probably right, the most important factor for the next Kepler is the TDP limitation. I actually doubt if those next gpus will be revolutionary, so it’s just a matter of putting more cudas on comparable dies. There’s also a ground for manipulation with the memory itself, it could provide a better performance by simply putting 320/384 bit interface on board.
Nevertheless, probably the most important thing that I expect from the next flagship graphics card is a better/quieter cooler. It’s time to make a competitive solution to those custom made by manufacturers. Imagine how could that affect the sales the reference models, which are usually cheaper.
BestJinjo
I don’t see how they can make a quieter solution though. I remember 7900GTX had an excellent cooler but ever since NV went for rear exhausting fans.
http://www.3dnews.ru/documents/12278/7900GTX-front-big.jpg
If they are going to increase TDP, I can’t see how it would be quieter than reference GTX680. With 250W TDP of 480, it was a loud card. I personally prefer an open-air design like your Gigabyte Windforce. I think the refernece blower works better in cramped cases, with poor airflow or in SLI/Tri-SLI. For single GPUs, I much prefer the cooler and quieter open air solution.
I am still split though. If next year, there are only 2-3 GPU demanding games (Crysis 3 and Metro Last Light), I am probably skipping that HD8000/GTX700. Right now preliminary benches for Crysis 3 show GTX690 getting crushed, so even a 50% faster GTX780 isn’t going to be enough. Based on the specs, Maxwell should bring a really large performance increase in 2014, just 1 year after Kepler refresh. I think that’s the one worth waiting for. Perhaps by then, there will be a lot more next gen PC games alongside DX11 PS4/Xbox 720. :)
http://videocardz.com/ WhyCry (VideoCardz.com)
You don’t see how they can make a quieter solution while mentioning the WindForce cooler. That’s exactly what they should do. How can a DirectCU or Windforce provide better airflow, more heat dissipation and yet be quieter, while costing almost the same amount of money (especially WindForce).
MOSLER
I think GTX 700-Series will be approximately so:
1) GTX 780 – 1920 CUDA cores (10 of 10 SMXs, each has 192 CUDAs), 160 TMUs (16 per SMX), 48 ROPs (6 of 6 raster engines enabled, 8 ROPs per engine) and 384-bit bus width (6 of 6 64-bit Memory Controllers [MCs] enabled). Possible clocks (GPU/Memory/Boost): ≈1050 MHz/≈1500 MHz/≈1100-1150 MHz.
2) GTX 770 – 1728 CUDA cores (9 of 10 SMXs), 144 TMUs, 48 ROPs (5 of 6 raster engines enabled) and 320-bit bus width (5 of 6 64-bit MCs enabled). Possible clocks (GPU/Memory/Boost): ≈925-950 MHz/≈1500 MHz/≈1050-1100 MHz.
3) GTX 760 Ti – 1536 CUDA cores (8 of 10 SMXs), 128 TMUs, 32 ROPs (4 of 6 raster engines enabled) and 256-bit bus width (4 of 6 64-bit MCs enabled). Possible clocks (GPU/Memory/Boost): ≈850-900 MHz/≈1500 MHz≈/≈950-1050 MHz.
4) GTX 760 – 1344 (or 1152) CUDA cores (7 or 6 SMXs), 112 or 96 TMUs, 32 ROPs (4 raster engines) and 256-bit bus width (4 64-bit MCs). Possible clocks (GPU/Memory/Boost): ≈800-850 MHz/≈1250 MHz/≈900-1000 MHz.
5) GTX 750 Ti – 960 CUDA cores (5 SMXs), 80 TMUs, 24 ROPs (3 raster engines) and 192-bit bus width (3 64-bit MCs). Possible clocks (GPU/Memory/Boost): ≈900-950 MHz/≈1375 MHz/≈1000-1050 MHz.
6) GTX 750 – 576 CUDA cores (3 SMXs), 48 TMUs, 16 ROPs (2 raster engines) and 128-bit bus width (2 64-bit MCs). Possible clocks (GPU/Memory/Boost): ≈850-900 MHz/≈1250 MHz/≈900-950 MHz or no boost.
What do you think? Is it close to possible specs?
It will be great if it be so!
Pingback: New NVIDIA Roadmap Reveals Maxwell Coming in 2014, GeForce GTX 780 Details - Quentin Rademaker Dot NL
BestJinjo
Ya, that’s a good point. I know they can make the cards much quieter if they go with an open-air design. I think they should. I don’t know for sure why they stuck with a blower fan for GTX480/580s. I think it was more to appeal to OEMs.
As far as Crysis 3 goes, I am hoping the performance will improve dramatically. Right now, it hammers GPUs on very high quality. I mean GTX690 can’t even get 45 fps at 1080P maxed out. I think mining will be completely dead soon. I don’t think mining will be a factor for me next generation! The rewards are getting halved next month and if specialized mining ASICs ship, GPU mining is dead as early as January. By next generation, I am fully expecting mining to be a non-factor.
I wonder if NV will try to sneak in a launch of GTX700 series right around Crysis 3 for marketing reasons. Bioshock Infinite, Metro Last Light, Starcraft 2: HotS, Company of Heroes 2 are also slated to launch around end of Feb-March 2013. It would be a very smart month to launch a new generation as February-March 2013 has a ton of games coming out.
Right now the rumors on GTX780 are all over the place, from 15-30%, but if GK110 is in play, it could be easily 40-50%.
BestJinjo
GK104 –> I assumed GTX780 will be based on similar ratios as I think there is a chance they might not use GK110.
(I) The ratio of GPC/ROP blocks-to-SMX clusters is 1:2 on GK104.
(II) The ratio of GPC blocks to TMUs is 1:32
(III) 1 GPC block = 2 SMX clusters and each SMX cluster = 192 CUDA cores. So 1 GPC block = 384 CUDA cores and 32 TMUs
GK104 has 4 ROP/GPC blocks with 8 ROPs / block. To get 384-bit bus you mentioned, you need 6 ROP/GPC blocks with 8 ROPs / each = 48 ROPs. If you have 6 ROP blocks, you have 6 GPC blocks and each GPC block has 384 CUDA cores.
You end up with 6 GPC blocks = 6 ROP blocks = 6 memory controllers @ 64-bit per block = 384-bit bus
6 ROP blocks @ 8 ROPs/bock = 48 ROPs
6 GPC blocks @ 384 CUDA cores / GPC block = 2304 CUDA cores
6 GPC blocks @ 32 TMUs / GPC block = 192 TMUs
That tells me if GTX780 is a derivative of GK104, but a larger version, you have 2 cases:
GTX780: 6 GPC blocks, 6 ROP blocks, 12 SMX clusters, 6x 64 memory controllers = 384-bit card, 3GB VRAM (because it’s 256 MB per 1 SMX cluster)
2304 SPs (12 SMX)
48 ROPs
192 TMUs
384-bit (6Ghz GDDR5 => 288 GB/sec)
3072 MB
GTX770: Would could have 1 GPC block removed. We get:
1920 SPs (10 SMX)
40 ROPs
160 TMUs
320-bit (6Ghz GDDR5 => 240 GB/sec)
2560 MB
Clocks around 950-1000mhz.
* Just speculating * If this is based on GK110, it would look different! My specs would mean GTX780 is much faster than 770 and it would allow NV to set a higher premium for the fastest single-GPU next round, beating HD8970, maybe $599 mark, or $649 depending on 8970′s price.
MOSLER
GPUs aren’t not based on any ratios. Developers and engineers decide how many cores, ROPs, TMUs etc. will be in GPU and they can change it as they wish. And these ratios are observed only in the cases of top- and high-end cards. GTX…70 and GTX…60 can has another noninteger ratios.
And GK110 probably won’t be used in gaming cards because it’s too expensive in production, not designed for overclocking and just wasn’t designed for mainstream. It’s only for Tesla cards. And if they use GK110 in GTX 780 it’s price will be about 800$. Just imagine how high will be the price of GTX 790 and ASUS Mars 4. It’ll reach $2000 price tag! It’s too expensive. It’s beyond reason.
And I don’t think they’ll disable 2 SMXs. They usually disable only one SMX, one 64-bit memory controller and down GPU clock.
My suggestion looks more objective, but your suggestion may also occur.
ZoSo
Jinjo, I wasn’t saying they didn’t have trouble 6 months ago. However he was asking about now and I posted ref article about GK110 yields now. That is all I was saying.
Also like on most of your posts, I agree with you about the GK110 and GeForce.
BestJinjo
GK104 and GK110 are built using ratios. That’s how all GPUs are built. All GK104 chips are built using the same ratios, except GTX660Ti has a wonky memory controller workaround. But if you don’t follow any ratio guidelines, you are just guessing to the specs:
http://www.anandtech.com/show/5699/nvidia-geforce-gtx-680-review/2
I am not saying your suggestions are wrong. I just built up mine using the same ratios as GK104 chip has, assuming NV will make a larger version of GK104. The alternative could be just using a cut-down GK110, but we already know the specs of that. At this point, since we don’t know if GTX780 will be built on an evolution of GK104 or an evolution of GK110, it’s almost impossible to guess the specs. Ya, I agree that disabling 2 SMX seems too much, which is why I have an alternative case for the GTX770.
I am leaning to agree with you that it would be more cost effective for NV to make a larger gaming chip than GK104 and GK110 is working backwards – it’s taking a huge bloated compute chip and removing parts to make it a gaming chip. I also think there is the issue of being able to manufacture enough GK110s to meet corporate and consumer demand. If they make a GK104 evolution chip, they don’t have to fight for allocating GK110 to corporate vs. consumer markets.
BestJinjo
Ya, gotcha; the yields are fine now. I think NV said so publicly. :)
My guess is voltage will still be unlocked for GTX700 series though.
MOSLER
I corrected GTX 770′s ROPs count. I accidentally wrote 48 instead of 40. I don’t throwing random numbers.
BestJinjo
Sorry, I didn’t mean to say you threw random numbers. I know you did a very careful analysis but you did use ratios just probably in your mind without thinking about them on the spot :)
modmax
gtx 780 = 2xgk106 960×2 192×2 = 200 eur x 2 ? NOOOOOOO!! 599 min.
kipdokken
if it is a refresh, i think the gtx 680 now will become gtx 770 in terms of performance this coming march. this is just my opinion.
svc
best of luck nvidia !

Club3D Announces Gamebundles For GeForce and Radeon Graphics Cards NVIDIA Unveils Tesla K20X

VideoCardz.com

New NVIDIA Roadmap Reveals Maxwell Coming in 2014, GeForce GTX 780 Details

GeForce GTX 700 Series – Kepler Refresh

The Kepler Refresh Naming

GeForce GTX 780 Launch Date

Follow us

Tag Cloud