[VideoCardz] NVIDIA to preview 20nm high-end Maxwell tomorrow

[Fastx]

Update 3-25-14 UPDATE: Apparently the list originated at NeoGraf forums, where it was clearly marked as ‘hoping to see’. Nothing official at this point.

It looks like we are going to hear a lot about Maxwell during Jen-Hsun’s keynote at GTC 2014.

Someone has just posted a short overview of the upcoming keynote, which would include a huge news to many GPU enthusiasts.

GeForce 800 series — 20nm Maxwell.

Jen-Hsun is going to preview 20nm high-end Maxwell GPU architecture. This is the first time we are going to officially hear about 20nm GPUs from NVIDIA. It also seems that second generation Maxwell is going to be used in GeForce 800 series. This is most likely just an overview of the architecture, something like the first showcase of GK110. We do not expect to see any actual samples.

DirectX 12, OpenGL and new driver

NVIDIA CEO is also going to talk more about the new DirectX and OpenGL APIs. Because the presentation is going to take around 2 hours, we expect many real-time demos, videos and people explaining those APIs in detail. New DX12 and OpenGL are a direct response to AMD’s Mantle. The new driver is supposedly going to introduce a huge difference in performance.

GeForce GTX 790 tomorrow?

Since it’s Jen-Hsun Huang, who is going to make the presentation, we are more than sure that he is going to show some new tech. Long story short, there is a good chance we are going to see GeForce GTX 790 tomorrow. Be advised, this is just my speculation.

Don’t forget to bookmark this link:

Live stream link


March 25th, 9:00 AM

It’s time again for NVIDIA CEO Jen-Hsun Huang to open the annual GPU Tech Conference with a 2-hour keynote. The show’s on from 9:00am to 10:50am.
•Preview of the new 20nm high-end Maxwell GPU architecture (GM10x/GeForce 8xx) with unified virtual memory.
•Amazing real-time graphics demos (as always).
•More DirectX 12 goodness.
•OpenGL goodness.
•Demos of the to-be-released CPU-optimized GeForce driver: http://www.pcper.com/reviews/Graphic…y-Improvements
•GameWorks demos.
•Exciting announcements.

Live Stream

http://www.twitch.tv/nvidia

Repository for GTC stuff and more: http://nvidianews.nvidia.com/imageli…&SubjectID=180

Related links:

http://www.gputechconf.com
http://www.geforce.com
http://blogs.nvidia.com
http://nvidianews.nvidia.com

http://videocardz.com/50012/nvidia-preview-20nm-high-end-maxwell-tomorrow

 

[T_Yamamoto]

It'll be interesting how it would compare with AMD for mining.

 

[OCGuy]

This will be the most highly anticipated GPU flagship release of all time if true, due to the overwhelmingly positive ink 750ti has received.

I am on the record as guessing GM2xx will hit the shelves before the standard narrative suggests. I guess we will find out! I still find it hard to believe they would have been able to keep such news under wraps.

It isn't a stretch to say that they could be alone @ 20nm for 6+ months.

 

[Gloomy]

If chips had left the foundry we would have gotten a leak already tho

 

[rtsurfer]

This will be the most highly anticipated GPU flagship release of all time if true, due to the overwhelmingly positive ink 750ti has received.

I am on the record as guessing GM2xx will hit the shelves before the standard narrative suggests. I guess we will find out! I still find it hard to believe they would have been able to keep such news under wraps.

It isn't a stretch to say that they could be alone @ 20nm for 6+ months.

I can am pretty sure that the flagship Maxwell wouldn't be out for 6+ months.

What would be interesting is how the GK104 equivalent cards of Maxwell generation compare to 780/780ti.

 

[Silverforce11]

It isn't a stretch to say that they could be alone @ 20nm for 6+ months.

It's a stretch because so many major players demand from so little production at TSMC. All the recent SKUs being released should clearly indicate we are still a long way from 20nm GPUs en-mass.

Not to mention, as soon as 20nm big-die Maxwell is available, imagine the backlog on Tesla orders to be filled first.

 

[Blue_Max]

Considering how good the GTX 750 TI is at mining/computing, I expect good things!

Something a little faster than the 750TI, plus SLI capable, plus same low-power goodness? Count me in!!

 

[el etro]

A Titan2 paper launch in early April? Not so good idea.

 

[escrow4]

Paid just under a $1K for a 780Ti GHz here recently, and I'll pay another $1K if the Maxwell 880 Ti is at least 30% faster . . . . . . .

 

[tviceman]

Flagship Maxwell die (GM100 / GM110) will overwhelmingly more-than-likely follow the release of every other significant 20nm maxwell die (GM104, GM106, and GM107). Kepler's first two chips to be released on 28nm were GK107 (first, in notebooks) followed shortly by GK104 (as the gtx 680). Nvidia learned their lessons from both trying to create smaller dies out of bigger ones (GT200 / Tesla) and leading with the large die first with a modular design (GF100 / Fermi). They'll lead with smaller dies first to learn as much as they can from the node to minimize defects and problems for the large die.

IMO it's a foregone conclusion of what we can expect specs-wise out of GM104 and GM107 Maxwell on 20nm (GM100/110 and GM106 are still big question marks, IMO). TSMC has said 30% performance at similar power consumption levels at 28nm (not sure which 28nm process was being used for comparison and may be a best case scenario), and 1.9:1 transistor density increase (also probably best case). If Nvidia achieves 30% perf/watt improvement and 1.6:1 transistor density improvement, then GM107 @ 20nm should be faster than both gtx660 and hd7870 with a die size of around 92.5 m^2. This may be a conservative estimate on perf/watt though, Nvidia is focusing very hard on that metric and will probably favor squeezing out perf/watt even if there is a slight die penalty. Greater than 30% perf/watt improvement may be easily achievable if TSMC was comparing 20nm with 28 HPM (which Maxwell is NOT being manufactured on). I think 20nm GM107 will end up 35% more efficient than 28nm GM107 with a die size around 100-110mm^2.

Extrapolating Kepler's lineup (GK104 being 4x the cores 2x the memory bus) would suggest that a 20nm GM104 will be 2560 maxwell cores and approximately 2.7-3x faster than 20nm GM107, making it about 35% faster than a gtx 780 TI.


TL;DR

I think Maxwell on 20nm will be comparable to Kepler in much the same way Kepler was comparable to Fermi. I think the performance jumps between architectures will be extremely similar.

 

[HurleyBird]

Can't wait, hopefully these are coming out soon. Might be a good thing the mining tax stopped me from buying a 290X!

 

[DarkKnightDude]

Can't wait for the sky high prices for 20nm. Going to be a great year.

 

[wand3r3r]

Great to hear something but I bet it will be as little info as they can reveal and only to keep people guessing.

This will be the most highly anticipated GPU flagship release of all time if true, due to the overwhelmingly positive ink 750ti has received.

I am on the record as guessing GM2xx will hit the shelves before the standard narrative suggests. I guess we will find out! I still find it hard to believe they would have been able to keep such news under wraps.

It isn't a stretch to say that they could be alone @ 20nm for 6+ months.

Remind me again what was the overwhelmingly positive ink the 750 ti has received besides for the power consumption.

Alone for 6+ months? Surely you jest.

Can't wait for the sky high prices for 20nm. Going to be a great year.

For stockholders.

 

[rtsurfer]

Can't wait, hopefully these are coming out soon. Might be a good thing the mining tax stopped me from buying a 290X!

Oh stop beating the dead horse.
The prices have come down to slight above the original & they frequently go to original retail levels but sell out quickly if you aren't looking out for them.

 

[OCGuy]

Flagship Maxwell die (GM100 / GM110) will overwhelmingly more-than-likely follow the release of every other significant 20nm maxwell die (GM104, GM106, and GM107). Kepler's first two chips to be released on 28nm were GK107 (first, in notebooks) followed shortly by GK104 (as the gtx 680). Nvidia learned their lessons from both trying to create smaller dies out of bigger ones (GT200 / Tesla) and leading with the large die first with a modular design (GF100 / Fermi). They'll lead with smaller dies first to learn as much as they can from the node to minimize defects and problems for the large die.

IMO it's a foregone conclusion of what we can expect specs-wise out of GM104 and GM107 Maxwell on 20nm (GM100/110 and GM106 are still big question marks, IMO). TSMC has said 30% performance at similar power consumption levels at 28nm (not sure which 28nm process was being used for comparison and may be a best case scenario), and 1.9:1 transistor density increase (also probably best case). If Nvidia achieves 30% perf/watt improvement and 1.6:1 transistor density improvement, then GM107 @ 20nm should be faster than both gtx660 and hd7870 with a die size of around 92.5 m^2. This may be a conservative estimate on perf/watt though, Nvidia is focusing very hard on that metric and will probably favor squeezing out perf/watt even if there is a slight die penalty. Greater than 30% perf/watt improvement may be easily achievable if TSMC was comparing 20nm with 28 HPM (which Maxwell is NOT being manufactured on). I think 20nm GM107 will end up 35% more efficient than 28nm GM107 with a die size around 100-110mm^2.

Extrapolating Kepler's lineup (GK104 being 4x the cores 2x the memory bus) would suggest that a 20nm GM104 will be 2560 maxwell cores and approximately 2.7-3x faster than 20nm GM107, making it about 35% faster than a gtx 780 TI.


TL;DR

I think Maxwell on 20nm will be comparable to Kepler in much the same way Kepler was comparable to Fermi. I think the performance jumps between architectures will be extremely similar.

Just a heads up to save you some key-strokes, the writers doing educated speculation Maxwell use GM2xx to denote the smaller node, be it 20nm or 16nm. Could be a fabricated naming convention, but it would serve a functional purpose.

 

[3DVagabond]

Can't wait for the sky high prices for 20nm. Going to be a great year.

Yeap! :thumbsup:

 

[OCGuy]

Not to mention, as soon as 20nm big-die Maxwell is available, imagine the backlog on Tesla orders to be filled first.

You are right about that. All the original Big K chips went to the Titan supercomputer. With the projected efficiency gains based on current Maxwell hardware, you can bet at least one national lab will design a Maxwell-based giant to continue our game of leap-frog with China. Supply would not be abundant, and I can see a release schedule exactly like Kepler's.

Nvidia has had Maxwell at the ready for years. They were given priority access to 28nm wafers @ TSMC, and secured early orders for 20nm before Apple was in the picture.

Does any of that mean I could be asking Santa for a 20nm Maxwell card? Not necessarily. But my guess is yes.

 

[OCGuy]

I think Maxwell on 20nm will be comparable to Kepler in much the same way Kepler was comparable to Fermi. I think the performance jumps between architectures will be extremely similar.

The main problem I see with this is that Maxwell is not all about more cores and improved efficiency.

Unified Virtual Memory is the one big ones that is certain. But there are rumours of more GPU innovations as well.

The next 3 days should be fun.

 

[Tristor]

If we see similar perf/watt gains in GM210 that we saw in GM107 with the launch of the GTX 750 Ti, I am going to be absolutely stoked to trade up my 780s for a couple of 880 (Ti)s for 4K Project CARS. I just hope they can meet timelines, because I really want to build an X99 beast Q1/Q2 of 2015.

 

[dangerman1337]

That text awfully looks similar to...

http://www.neogaf.com/forum/showthread.php?t=790013

Hoping To See...

- Preview of the new 20nm high-end Maxwell GPU architecture (GM10x/GeForce 8xx) with unified virtual memory.
- Amazing real-time graphics demos (as always).
- More DirectX 12 goodness.
- OpenGL goodness.
- Demos of the to-be-released CPU-optimized GeForce driver: http://www.pcper.com/reviews/Graphic...y-Improvements
- GameWorks demos.
- Exciting announcements.

Looks like someone just copy pasta non-insider post from NeoGAF about someone's hopes about what is too be revealed and has been passed around as legit info. Also there isn't a GM100/104, only GM200, GM204 & GM206 which the latter three have cropped up as codenames before.

Regardless I hope they do give us Second generation Maxwell info and will come relatively soon but I'd keep my hopes low. I really want a 4GB 512-bit GM200 for 2560x1440 for the Witcher 3, Consumer Version of the Oculus Rift and games not designed to accommodate 7-8 year old hardware. Though I won't be surprised if Nvidia just crams in more CUDA/shader cores

 

[Mondozei]

Lets see if their livestream works better than the disaster we saw when AMD launched Hawaii. Im guessing at least as many will watch, if not more.

 

[Lonyo]

That text awfully looks similar to...

http://www.neogaf.com/forum/showthread.php?t=790013

Looks like someone just copy pasta non-insider post from NeoGAF about someone's hopes about what is too be revealed and has been passed around as legit info. Also there isn't a GM100/104, only GM200, GM204 & GM206 which the latter three have cropped up as codenames before.

Regardless I hope they do give us Second generation Maxwell info and will come relatively soon but I'd keep my hopes low. I really want a 4GB 512-bit GM200 for 2560x1440 for the Witcher 3, Consumer Version of the Oculus Rift and games not designed to accommodate 7-8 year old hardware. Though I won't be surprised if Nvidia just crams in more CUDA/shader cores

They updated the link the OP sourced his quote from to note that they did indeed just copy from NeoGAF.
So nothing to speculate on really, just a wait and see what happens situation.

 

[OCGuy]

WCCFtech (or whatever it is) regurgitating a blog post by Videocardzzz!!11 who copied it from whoever.

That makes Charlie look like John Stossell.


Anyway keynote is coming up in the next couple hours, so hopefully we know more by the end of the day.

 

[Zanovar]

Looking forward to more info:thumbsup:

 

[Slomo4shO]

The stream is live.