Nvidia Pascal GPU has 17 billion transistors

[BoFox]

NVIDIA GeForce GTX 1080 DirectX 12 Benchmarks in Ashes of The Singularity Revealed – Async Compute Performance Analyzed
http://wccftech.com/nvidia-geforce-gtx-1...enchmarks/

HUH??

[RolloTheGreat]

NVIDIA GeForce GTX 1080 DirectX 12 Benchmarks in Ashes of The Singularity Revealed – Async Compute Performance Analyzed


http://wccftech.com/nvidia-geforce-gtx-1...enchmarks/

HUH??

Quoted to link the article.

I'm not a fan of the "Ashes of AMD" benchmark, but why "HUH??"?

[SteelCrysis]

They also say that more reviews of Pascal are coming on May 17.

[ocre]

Okay Ocre, in no way am I saying nVidia should be doing things this way, just going to lay out an alternative path they *could* have taken which I would have found more agreeable.

GP100 is already in production, based on the information we had it was prior to GP104. Now, if they had a fully GPU functional GP100 base and started binning them from day one and launched in say August instead of May, they could have saved all the say 1/4 defective cores for Titans, 1/3 defective for 1080s, 1/2 defective for 1070s. If they had done that and then launched the GK104 parts as 1060Ti and 1060 we could be looking at this lineup-

Titan- $1K
1080- $650
1070- $450
1060Ti- $350
1060- $270

Now the x60 parts would have less and slower RAM so they wouldn't be as fast, but we'd still be looking at an overall *significantly* faster top to bottom lineup. 

Don't get it confused in any way, I completely understand *why* they do it the way they do, just saying as an enthusiast I'd rather see the big boy sooner. For the record, I had the exact same issues when the 680 launched, which I why I waited for the 780. 

If they launched with that lineup, AMD's new architecture would be looking at their highest end part fighting for fifth fastest GPU, and honestly, that would likely shut the viral marketers up which to me would be the best possible thing that could happen to enthusiast forums.

I don't think the volume capability is there at all. This massive size chip has yield and Hbm2 working against it.
I am not even sure how well capacity will be with the gp104 but I have a feeling it will sell out and be awhile before things level out.

I mean, there is no way to know for sure. But it only seems rational. If the big chip is all ready , there is no reason to hold it back. They could have the King Titan launch at 1000-1200 first and it would do absolutely nothing to hurt nvidia in gp104 sales.

They would sell far more gp104 chips if they were priced less, this you know. As does nvidia. But there is much more to consider, like capacity and how many you can sell. Seems logical that gddr5x will not be endless supply right off the bat. But again, it's speculation.

I will say this much, if nvidia had the ability to launch the gp100 in volume---there is no reason not to. The price of King titan when it comes will be a lot higher that the 1080. They would have had so much more to brag on in their pascal presentation. So so much more.

I believe the big chip is in production, it's gonna be slow roll out. That the Tesla markets are dieing for them. See, if yields are only 25%, it might still be cost worthy for them to sell chips from low yield wafers, but you better believe that nvidia will be wanting to increase yields significantly before they go to full production.

And this is besides the fact that hdm2 is beyond their control. It will have to be ramped up as well

[RolloTheGreat]

And...could be they're launching multiple times to get high end users starved for parts to buy twice.

It would make a LOT of sense from a business perspective to do this, especially in the absence of competition.

[BoFox]

NVIDIA GeForce GTX 1080 DirectX 12 Benchmarks in Ashes of The Singularity Revealed – Async Compute Performance Analyzed


http://wccftech.com/nvidia-geforce-gtx-1...enchmarks/

HUH??

Quoted to link the article.

I'm not a fan of the "Ashes of AMD" benchmark, but why "HUH??"?

Oh, thanks - must've forgotten to use the URL insert..

I'm just confused by the performance, since it seems that it's not really any faster at Async - more than the generic speed increase over the 980 Ti.

[BoFox]

Wow, it's been so long ago (28nm being so long in the tooth) that I actually pretty much forgot about this, but now I remember thanks to you!

AMD's move is just pathetic - AMD should have had a hidden Ace up its sleeve, ready to trump the game after NV launched GTX 1080.  Nvidia already knows that it's like 8800GT/9800GTX vs HD 3870 all over again, with AMD being careful on new 14nm GloFo process. 

I think that AMD desperately wants to beat NV to the punch with HBM2 memory, so AMD is probably placing all bets on this waiting game (early 2017).

Rumor posted 11 minutes ago over at VCZ:
AMD allegedly pushes VEGA GPU forward to October

Perhaps not so pathetic of AMD after all - beating NV to the punch with HBM2!

[ocre]

Bofox, according to rollo's theory, nvidia is/may just be holding back the gp100 to starve high end users. So amd isn't beating nvidia to the punch, nvidia is just starving everyone because it makes business sense.....

Starving Tesla market because it makes business sense. The 1080 will kill off titan x sales and not having an ultra expensive pascal titan to replace it, makes business sense....somehow.

I, on the other hand, think that nvidia couldn't be in anymore of a hurry to get gp100 Tesla cards on the market. But, I could be wrong.

As for the new rumor you just posted, I actually have been expecting something like this to "leak" out. Obstruct when you can't compete, that has been in AMD'S play book for some time now. You have witnessed it, although you may not have put it together yet. There are plenty of examples, freesync should be fresh on your mind. AMD did all they could to obstruct and hold off people from going gysnc. They were no where close to a competitive solution but had ever one believing that gysnc was nothing but a money grab and they would have a better solution just around the corner....which of course, they knew wasn't true. There was no monitors and it was a long long way off, but yeah...just around the corner turned out to be an extremely long time. And freesync, well it wasn't free either.

Anyway, the timing of this is all you need. Some random guy post something and you better believe that it's gonna be spread throughout the entire interwebs by rumor articles and amd super troopers. "Vega is just around the corner, only 3-4 months away. It's gonna be faster and cheaper and more efficient. It will beat pascal on every metric, by a lot. It's dumb to buy a card today, you should at least wait to see what amd has up their sleeve with vega, it's just a tiny wait". You can imagine the drill.

See, it doesn't matter if vega launches in Oct or not because if they can get people to wait that long then it is just rinse and repeat. Some made up benchmarks and new planted info, then you go at it again. OCTOBER- see look at the charts, vega I'd amassing!!! It's just around the corner. If people start questioning...they will point to the unsubstantiated alleged vega performance figures and say that this is proof that it is just around the corner. And since the charts will be most likely made up, they should look pretty darn good. Anyone who had waited and hoped for that long, they will easily eat it up. See, this will make them feel validation, even better about waiting.
No matter how far vega is away or where it will per, amd would much rather stop people from buying nvidia cards. Obstruct

[RolloTheGreat]

I'm just saying it's possible and why it could make sense.

NV releases mid range part at $600-$700 that all enthusiasts buy because there is nothing else.

They sell big chip which has lower yields at first to pro market for big bucks.

After a few months of that chips with failed clusters of processors are sold as the 1080Ti for $999..

I agree on the HBM2 thing though so you could be right.

[SteelCrysis]

GDDR5X is in production ahead of schedule: https://www.micron.com/about/blogs/2016/...ron-gddr5x

[SteelCrysis]

Nvidia is reportedly focusing on 2-way SLI with Pascal: http://www.techpowerup.com/222426/nvidia...rce-pascal

Edit: Look at Zotac's GTX 1080 PGF! http://www.techpowerup.com/222430/zotac-...phics-card

[BoFox]

Me buying one!

[ocre]

Designed by aliens!!!

[RolloTheGreat]

Me buying one!

That is a likely purchase for me as well. Rollo Jr is all about the case bling.

[SteelCrysis]

Final specifications for Pascal: http://videocardz.com/59962/nvidia-gefor...esentation

[SteelCrysis]

Review is in: http://www.tomshardware.com/reviews/nvid...72-12.html

Enabled in part by TSMC’s 16nm FinFET+ process, the 1080’s GP104 GPU wields 2560 CUDA cores at an unprecedented 1607MHz base clock rate. It offers a substantial step up from GM204 and an impressive boost compared to the former flagship GeForce GTX 980 Ti. In fact, across the eight real-world games we benchmarked today, GeForce GTX 1080 averages 34%-higher frame rates than the 980 Ti at 3840x2160. That’s enough performance to let you run with taxing detail settings and still enjoy the experience. Presumably, it’ll make an equally potent (if not better) VR solution, particularly as software developers get their hands on the Pascal architecture and start exposing features like Simultaneous Multi-Projection.

[SteelCrysis]

http://www.pcgamer.com/gtx-1080-review/

With manual tuning, we settled on a +200MHz core overclock, +725MHz on the GDDR5X (11.45GHz), with the overvoltage set to "100%" (whatever that means—it's not a doubling of the voltage, just to be clear). In practice, looking at actual GPU Boost 3.0 clock speeds running stock vs. overclocked, we saw 2.0-2.1GHz on the core (vs. 1700-1733MHz), so we got a healthy 15 percent overclock on the core and memory. Results may vary, but our sample card looks to be a bit less impressive than other cards, so 2.0GHz and more looks attainable.
...
Performance improves on average by 10-15 percent, except at 1080p where we start to hit CPU limits in a few titles. We'll likely see factory overclocked cards with aftermarket coolers at some point, and some of those will almost certainly exceed what we're able to achieve with the Founders Edition. Even better, many of the factory overclocked cards will likely cost less than the Founders Edition. But if the 30 percent jump in performance over Titan X wasn't enough to get you excited, perhaps tacking on another 15 percent with overclocking will do the trick.

[SteelCrysis]

[SteelCrysis]

GTX 1070 specs revealed: http://www.techpowerup.com/222537/nvidia...s-revealed

[SteelCrysis]

GTX 1080 hit only 67 C at Nvidia's revealing of it because it was in GPU rendering, rather than gaming:

[SteelCrysis]

Nvidia will still partially support 3-way and 4-way SLI by allowing developers to support it, 3-way and 4-way SLI are still possible through third-party bridges and Nvidia Enthusiast Keys:

[SteelCrysis]

GTX 1070 clock speeds revealed as 1506 MHz base/1683 MHz boost, and 2000 MHz VRAM: http://www.techpowerup.com/222570/nvidia...s-revealed

[SteelCrysis]

Poll on whether people will be upgrading to the GTX 1080: http://techreport.com/news/30153/poll-wi...e-gtx-1080

[ocre]

There is a lot of haters out there. Amd troopers in full force. No place is safe...

It's all in vain though.
Polaris looks like it will be a joke compared to pascal. Can't wait to see how this all plays out

[BoFox]

Wow, the pre-launch rumors were stating GTX 1080 to be about 20% faster than Titan X, but now the reviews are actually showing more than 25% overall.  That is quite pleasant.  

I'm eagerly awaiting custom versions with 2 PCI-E power plugs that would allow more power for higher overclockability.  I want to be promised at least 2.1 GHz.  2 GHz does sound nice nonetheless.  

It's actually amazing how a chip barely more than half the size of Titan X (just barely, barely more than 1/2 the size) is actually considerably faster.  Kudos to Nvidia for pulling this off, without resorting to HBM2 memory.

4K is becoming mainstream pretty soon!

[BoFox]

Wow, I'm finding GTX 1080 to be extremely bandwidth-starved in most of the current demanding games:

http://www.computerbase.de/2016-05/gefor...r-gtx-1080
http://www.computerbase.de/2016-05/gefor...ark-part-1
http://www.computerbase.de/2016-05/gefor...ark-part-2
http://www.computerbase.de/2016-05/gefor...ark-part-3

[RolloTheGreat]

There is a lot of haters out there.  Amd troopers in full force.  No place is safe...

It's all in vain though.
Polaris looks like it will be a joke compared to pascal. Can't wait to see how this all plays out

AMD is fighting for survival the only way they can- a guerrilla war of FUD.

It won't help them, but it's all they have.

[SickBeast]

Hey so going back to the OP, I'm guessing that the 17 billion transistor count is for big Pascal? The GTX 1080 has nowhere near that much, I don't even think it has half that. Perhaps big Pascal is the GPU that we should all be really excited about.

[RolloTheGreat]

Hey so going back to the OP, I'm guessing that the 17 billion transistor count is for big Pascal?  The GTX 1080 has nowhere near that much, I don't even think it has half that.  Perhaps big Pascal is the GPU that we should all be really excited about.

I'm pretty excited about what the 1070/1080s bring to market.

1080s are the most legitimate single GPU for 4K to date.

[SickBeast]

Hey so going back to the OP, I'm guessing that the 17 billion transistor count is for big Pascal?  The GTX 1080 has nowhere near that much, I don't even think it has half that.  Perhaps big Pascal is the GPU that we should all be really excited about.

I'm pretty excited about what the 1070/1080s bring to market.

1080s are the most legitimate single GPU for 4K to date.

I'm kind of torn as to what to do. The GTX 1080 is clearly overkill for 1080p gaming. So that leaves me the GTX 1070 which I am kind of interested in. I was planning my next GPU upgrade around my transition to 4k. I just don't know what to do any more though. I would clearly enjoy the GTX 1070 for 1080p gaming until I upgrade my TV to 4k, however I want an HDR set and they are way too expensive right now.

Titan X performance is very tempting for me so I will be closely watching the Polaris launch. If it can compete with the GTX 1070 then perhaps we can have a price war and some good deals. Then I will be in for sure!

[BoFox]

Hey so going back to the OP, I'm guessing that the 17 billion transistor count is for big Pascal?  The GTX 1080 has nowhere near that much, I don't even think it has half that.  Perhaps big Pascal is the GPU that we should all be really excited about.

Nah, not that much - Tesla P100 (GP100) is already announced to have 15 billion trannies - a bit more than 2x that of GTX 1080.  Die size is roughly twice as big, as well.  It has HBM2 memory, but it's clocked at only 1400MHz and has the same 4096-bit total bus width, rather than 8192-bit as a few of the rumors were claiming.  It's not bad, compared to 1000MHz HBM on the Fury cards - what's for sure is that it will have 16GB of RAM - perhaps even as much as 32GB for a special edition.

[BoFox]

Wow, I'm finding GTX 1080 to be extremely bandwidth-starved in most of the current demanding games:

http://www.computerbase.de/2016-05/gefor...r-gtx-1080
http://www.computerbase.de/2016-05/gefor...ark-part-1
http://www.computerbase.de/2016-05/gefor...ark-part-2
http://www.computerbase.de/2016-05/gefor...ark-part-3

Are any of you guys surprised by this?  Even more aggressive bandwidth compression that Pascal brings isn't enough.  

The memory is overclocked by 10% (from 5000 to 5500MHz, or 11000MHz effective)..  while the core is left alone.

AC Syndicate is showing an 8% increase in performance.
Far Cry Primal shows a whopping 9% increase.
GTA V also shows 8% increase.
Heroes VIII - also 8%
The Division - 8%

15 games out of 21 tested show 5% or greater increase in performance.

All of this owing to a 10% increase in memory bandwidth ALONE????

This is incredibly hard to ignore.  

I'm hoping that the site didn't accidentally turn up the power limit from 100% to say, 120% or something like that?  Even the fan profile could've been accidentally tweaked - it's very easy to overlook one of these things, as overclocking utility programs usually save the fan profile which is usually automatically applied the next time the program is launched.  

Yet, I don't think ComputerBase is that dumb to overlook such things.  Perhaps another site will go over this, unless I missed one.  Reading Techreport now..

[RolloTheGreat]

Wow, I'm finding GTX 1080 to be extremely bandwidth-starved in most of the current demanding games:

http://www.computerbase.de/2016-05/gefor...r-gtx-1080
http://www.computerbase.de/2016-05/gefor...ark-part-1
http://www.computerbase.de/2016-05/gefor...ark-part-2
http://www.computerbase.de/2016-05/gefor...ark-part-3

Are any of you guys surprised by this?  Even more aggressive bandwidth compression that Pascal brings isn't enough.  

The memory is overclocked by 10% (from 5000 to 5500MHz, or 11000MHz effective)..  while the core is left alone.

AC Syndicate is showing an 8% increase in performance.
Far Cry Primal shows a whopping 9% increase.
GTA V also shows 8% increase.
Heroes VIII - also 8%
The Division - 8%

15 games out of 21 tested show 5% or greater increase in performance.

All of this owing to a 10% increase in memory bandwidth ALONE????

This is incredibly hard to ignore.  

I'm hoping that the site didn't accidentally turn up the power limit from 100% to say, 120% or something like that?  Even the fan profile could've been accidentally tweaked - it's very easy to overlook one of these things, as overclocking utility programs usually save the fan profile which is usually automatically applied the next time the program is launched.  

Yet, I don't think ComputerBase is that dumb to overlook such things.  Perhaps another site will go over this, unless I missed one.  Reading Techreport now..

https://www.techpowerup.com/reviews/NVID...80/16.html

Over 30% faster at Far Cry Primal at 4K, 40fps vs 30fps on a Titan X.

https://www.techpowerup.com/reviews/NVID...80/17.html

Same thing at GTAV, 45fps vs 35fps.

With our 4K GSync monitors, this difference is HUGE.

Memory bandwidth or no, this is the single GPU to have and my only question is whether I order Founders on Day 1 or can hold out for custom.

[SickBeast]

The problem is that big Pascal is going to crush it.

[RolloTheGreat]

The problem is that big Pascal is going to crush it.

That's kind of the nature of this video card thing, they're a fast depreciating asset. Computer gaming isn't for poor folk.


Always newer, faster cards coming out, supply and demand set price.

When will "big Pascal" crush it though? Lots of "big Pascals" or HBM2 being produced these days?

So if "big Pascal" comes out in February, was a 1080 for $600-$700 in May a bad deal? Methinks not. Can always be sold when new parts come out, and GHTX980s were selling for $350-$400 a couple weeks ago.

I got $520 for the 980Ti that cost me $650 a year ago.

[SickBeast]

Big Pascal will probably sell for $999USD, perhaps even more than that. I'm trying to remember what happened to GTX 680 prices once the Titan X came out.

It's definitely not a bad move to buy now if you need an upgrade. Vega and Big Pascal will probably just become halo cards and they will not affect the prices of the other cards.

[SteelCrysis]

GTX 1080M reportedly is faster than Titan X: http://videocardz.com/60211/nvidia-gefor...tx-titan-x

[SteelCrysis]

GTX 1080 preorders have begun on Amazon, and Newegg lists the GTX 1080 as coming soon.

[SteelCrysis]

HardOCP is defending Nvidia, confirms that aftermarket-cooled GTX 1080s will launch on the same day as Founder's Edition GTX 1080s: http://hardocp.com/article/2016/05/18/ge...aunch_ever

[SteelCrysis]

GTX 1080s sell out in minutes on Amazon, Newegg, and Best Buy: http://www.eteknix.com/nvidia-gtx-1080-p...n-minutes/