[SemiAccurate] Nvidia Pascal over a year ahead of 14/16nm competition

Status
Not open for further replies.

desprado

Golden Member
Jul 16, 2013
1,645
0
0
Nvidia is more than a year ahead of any competition in the mobile space as their Pascal based Drive PX 2 module proved. Held up by no less than CEO Jen-Hsun Huang, SemiAccurate felt it was an automotive tour de force.
The losses of Audi/VAG were a thing of the distant past at Nvidia because of the new Drive PX 2 module and specifically its twin 16nm Pascal GPUs. Teamed up with a total of eight ARM Cortex-A57 cores and four Nvidia Denver cores, full specs at Anandtech here. Together they provide everything a modern automobile could need for compute in a mere 250W package. That number is more than 10x the power draw of competitive solutions so it must be fast!
Getting back to their lead, Nvidia is so far ahead of the industry it is, well, nigh on incredible. Until the CES unveiling, AMD was widely believed to be 2+ quarters ahead of Nvidia on the process front, their 14nm Polaris GPUs had all the mindshare. On Monday January 4th, 2016, Nvidia showed undeniable proof that they were more than a year ahead of AMD in the race to 14/16nm devices. Not only that they were ahead of assumed leaders Apple and Samsung to devices on this node, way way ahead.
Why do we say this? Take a close look at the below picture from the official Nvidia Flickr stream of the Drive PX 2 board held up by Nvidia CEO Jen-Hsun Huang. This was taken at the CES unveiling of the PX 2, and you can see the full picture here. The room is full of press and analysts so SEC rules mandate that everything said is absolutely true, it would be illegal to do otherwise. With that in mind, you know this killer device is both real and what it is said to be.
The picture below is a closeup of one of the two Pascal GPUs on the PX 2 board that Jen-Hsun was holding. These 16nm FinFET devices are said to push out a combined 16TF of SP compute, a massive number for only 250W. That is only possible because of the process tech it is based on, 28nm devices would struggle to have half that performance in the same power envelope. But showing off 16nm pre-production parts is not a big deal, SemiAccurate held an AMD 14nm device in December. Why do we say that Nvidia has a massive lead over AMD then?


If you look closely at the chips you can see the date codes on them are 1503A1 and they are from TSMC. TSMC date codes are formatted as year of manufacture followed by work week of manufacture followed by stepping. For example the GTX 980 that was sent to the press in September of 2014 had a date code of 1421A1, the press GTX 960 from January 2015 had a date code of 1442A1, and the GTX 980Ti sampled in May of 2015 had a date code of 1436A1. You can work out the details of the press samples for yourself when you are bored.
What this means Nvidia has had 16nm FinFET based Pascal chips sitting on the shelf for almost a year, barely a quarter after Maxwell was produced. That silicon was manufactured in mid-January of 2015!!! Everyone including the author was assuming that Pascal didn’t even tape out until Q3 of 2015, we were off by at least nine months, more likely a year. More importantly this version of Pascal somehow doesn’t need HBM memory to reach those heady numbers, it uses vanilla GDDR5 chips as you can see.
That means the consumer versions of Pascal that have HBM will be so much faster than the quoted numbers for the Drive PX2 module that uses the GDDR5 memory that it is nigh unbelievable. Nvidia’s Pascal is so efficient that it can support compute levels that everyone believes is impossible with the memory bandwidths offered by the memory pictures. 16TF with GDDR5 is patently impossible for other silicon providers for anything but trivial benchmarks.
How Nvidia made the big Pascal chip in such a way that both GDDR5 and interposer based HBM memory will work with the same package is going to be very interesting to hear, every other company using HBM has found this impossible to implement too. More to the point such a memory configuration shouldn’t be possible using even bleeding edge technologies, but Nvidia has had this in their pocket, literally sitting on the shelf since January of 2015!!! That is technological leadership if it ever existed, quite the commanding lead.
But wait, the Nvidia leadership position is even more dominant than that. Those WW3/2015 date codes mean Nvidia had to tape their Pascal designs out in late 2014, even before the TSMC process was ready for such tapeouts much less stable. The Nvidia designers were obviously so good and so far ahead of the game that they were able to put out a design that worked so well out of the box that even on an unstable, some go so far as to say non-existent, process it worked on the A1 stepping. Not only that it worked so well that once TSMC stabilized the 16FF+ process months after Nvidia produced their Pascals, no update to the GPUs were even needed! Imagine that, so many critical PDK updates from TSMC and none managed to do any better than those plucky geniuses at Nvidia could get quarters before. That is a commanding lead if there ever was one, simply staggering process tech.
So to wrap it up, Nvidia CEO Jen-Hsun Huang showed off what he said were two Pascal GPUs in an Nvidia Drive PX 2 module. The date codes on the devices he held up made it very clear that they had been sitting on these devices for a week or two short of a year, no one else was even close to large 16nm FinFET GPUs at the time. Even TSMC could not produce such parts for months afterwards, their process wasn’t actually ready at the time. And the performance is more than 2x that of 28nm AMD devices even without the HBM memory that Pascal was supposed to require.
It all has to be true because as an executive in a publicly traded company with press and analysts who cover and trade Nvidia in the room, he is obliged by the SEC to be truthful. The only other explanation is that he knowingly lied to the press and analysts and showed a fake card in his keynote, something that would clearly be illegal.

http://semiaccurate.com/2016/01/11/nvidia-pascal-over-a-year-ahead-of-1416nm-competition/
I do not know he is right or wrong ,however, first time i have seen he begin positive about Nvidia.
 

Stuka87

Diamond Member
Dec 10, 2010
6,240
2,559
136
How can an unreleased product be a year ahead of an unreleased product? Or rather, how can anybody tell who is ahead of what?

EDIT: After fully reading the article, its very obvious he is poking fun at nVidia.
 
Last edited:

Hitman928

Diamond Member
Apr 15, 2012
5,182
7,632
136
I think you guys are missing the sarcasm in the article. He's making fun of Nvidia.
 

96Firebird

Diamond Member
Nov 8, 2010
5,709
316
126
I'm not sure if it is a good thing to be involved with VAG right now... I'm sure the whole cheating emissions scandal happened after Nvidia partnered with them, but it has already had an impact on VAG's reputation. Not to mention how much they'll have to pay out in the end.
 

NTMBK

Lifer
Nov 14, 2011
10,208
4,940
136
Full marks for reading comprehension, guys :rolleyes: He's making fun of NVidia for using a mockup (again). This isn't a positive article, at all.
 

Bryf50

Golden Member
Nov 11, 2006
1,429
51
91
Lol at the people in this thread that haven't realized this it's all sarcasm.
 

NTMBK

Lifer
Nov 14, 2011
10,208
4,940
136
How can an unreleased product be a year ahead of an unreleased product? Or rather, how can anybody tell who is ahead of what?

Did you read the article? The chip has a datecode stamped on it, telling us the week it was manufactured.

(Spoiler: It's a GTX 980M MXM card, standing in for Pascal because Pascal isn't ready to be shown off yet.)
 

xthetenth

Golden Member
Oct 14, 2014
1,800
529
106
OP is a joke, both the post and the poster.

I especially like the reference to the companies that have shipped silicon on the nodes being behind NV. That and the whole thing being a run up to either these miraculous, impossible achievements are true or he straight up lied on stage. Bringing in VW is also good to set the stage, it's a really nicely paced well done article.

It's better than an article saying in the Fermi run-up that NV has a head start on going green because their screws are ready for wood.

I also want to know who hacked the site because it's actually competent tech writing.
 
Last edited by a moderator:

nvgpu

Senior member
Sep 12, 2014
629
202
81
http://forums.anandtech.com/showpost.php?p=37443058&postcount=227

Charlie Demerjian said:
The GTX980Ti is nothing more than the high bin 980, this time it is clocked at an even 1GHz.

Everything else should be the same as the non-Ti 980s, it isn’t the big Maxwell the BS rumor sites listed a while ago, it is just what we said it was last year.S|A

This is all you need to know about Charlie Dermejian, absolute loser at life.
 

moonbogg

Lifer
Jan 8, 2011
10,635
3,095
136
I saw the word "mobile" and stopped reading.

"Nvidia is more than a year ahead of any competition in the mobile space..." - moonbogg made it this far.
 

Kenmitch

Diamond Member
Oct 10, 1999
8,505
2,248
136
JHH held up a board with Maxwell GPUs on it, and called it Pascal. So yeah, business as usual, I guess. *shrug*

Too funny!...If JHH says it's Pascal then it is....Rebadged is all.

I guess if you don't really have anything real to show then you just use a mocked up product and pretend it's working.
 

Glo.

Diamond Member
Apr 25, 2015
5,661
4,419
136
I have read the article when it was posted on TT.

Then I saw you posted here, and had good laugh.

First because of the article was quite good in making fun of Nvidia.

And then from people who did not saw the sarcasm in Charlie's writing.
 

AnandThenMan

Diamond Member
Nov 11, 2004
3,949
504
126
I do not know he is right or wrong ,however, first time i have seen he begin positive about Nvidia.

Not the first time, before Kepler came out he predicted the performance and power efficiency.

But this....
But wait, the Nvidia leadership position is even more dominant than that. Those WW3/2015 date codes mean Nvidia had to tape their Pascal designs out in late 2014, even before the TSMC process was ready for such tapeouts much less stable. The Nvidia designers were obviously so good and so far ahead of the game that they were able to put out a design that worked so well out of the box that even on an unstable, some go so far as to say non-existent, process it worked on the A1 stepping. Not only that it worked so well that once TSMC stabilized the 16FF+ process months after Nvidia produced their Pascals, no update to the GPUs were even needed!
Nvidia has really stepped up their game, incredible. :D But seriously how could someone read this article and not realize it was satire? It's actually pretty damn funny but man I don't know how anyone reading that could take it seriously.
 

Kenmitch

Diamond Member
Oct 10, 1999
8,505
2,248
136
Lisa Su held up Fury X and claimed it's overclockers dream. Business as usual.

So it's OK to hold up a fake product and call it real?

Not really sure why you'd compare the real Fury X to a faked mocked up product.

The overclockers dream comment in no way plays a role in real vs fake....Just in case you didn't realize it.
 

AnandThenMan

Diamond Member
Nov 11, 2004
3,949
504
126
Lisa Su held up Fury X and claimed it's overclockers dream. Business as usual.
Yea but without the woodscrews sawed off PCB and non-working silicon it was just wasn't the same. :'(

Anyway I'm sure Nvidia will be out with 16nm products very close to AMD on 14, this happens every generation/node change one supposedly has a massive lead etc. but it never happens.
 

Stuka87

Diamond Member
Dec 10, 2010
6,240
2,559
136
Yea but without the woodscrews sawed off PCB and non-working silicon it was just wasn't the same. :'(

Anyway I'm sure Nvidia will be out with 16nm products very close to AMD on 14, this happens every generation/node change one supposedly has a massive lead etc. but it never happens.

This is because they shared the same fab. With AMD now using Samsung, who knows what will happen to timing.
 

SlowSpyder

Lifer
Jan 12, 2005
17,305
1,001
126
Lisa Su held up Fury X and claimed it's overclockers dream. Business as usual.


I remember that, but thought that she was talking in the context of how the part is built and cooled with liquid cooled VRM's and the hybrid GPU/HBM cooler.


Charlie did have a very pro-Nvidia article out before the GTX680 launch, saying how good of a part it was at the time.
 
Last edited:
Status
Not open for further replies.