How good will Pascal be?

[poofyhairguy]

What I want is 280X worth of power with 750 ti power consumption levels (aka doesn't need a 6 pin connected).

 

[moonbogg]

Come on poof. That ain't happenin.

 

[Headfoot]

So much insight here. I am truly honored that you provided this unique and wonderful thought. What would the AT board be without your industry acumen?

Savage.

 

[moonbogg]

Savage.

And filthy. Don't forget filthy.

 

[Timmah!]

Yes. Guru3d has a rumor on their news page that the big Pascal GPU will be released in April this year. They always cost at least $1,000.00, but remember, the same cards used to cost about $500.00.

I hope this to be true, but i think at best they will just reveal GP100 Teslas in April on that upcoming conference (GTC?)...maybe start selling shortly after that. Consumer chips, god knows...

Just to remember, Kepler GK104 (GTX680) were released in March 2012, in April or May they unveiled the GK110 (aka Tesla K20) on the GTC, which becamse consumer product in February 2013 in form of the first Titan card. Which was later succeeded by GTX780Ti (November 2013) and later Titan Black (February 2014).

Not saying it will play out similarly with Pascal, but i would not be surprised if it did. At least partially. Although in regard to Titan, this is sort of different situation, before 2013 there was no Titan product at all, but since then, there was one every single year (the ones i mentioned above + Titan X in March 2015)...i wonder if they would had courage to release 2016 Titan build on GP104, if they save GP100 just for Teslas this year, or they rather would not release it this year at all...

....and then there is mysterious GP102...

 

[poofyhairguy]

Come on poof. That ain't happenin.

A man can dream.

 

[lehtv]

A man can dream.

Well, you may not get 280X level performance but you could get pretty close. When 750 Ti was released, it was 75% faster than GTX 650 which still required a 6-pin connector. HD 7750 was a few percent slower than GTX 650 but didn't require that 6-pin (at the time, Techpowerup called it "[the] most power efficient graphics card in the world").

R9 280X is about 100% faster than GTX 750 Ti, and Pascal should offer a pretty good jump forward in efficiency thanks to the new process node. So... if Pascal will be a bigger jump over Maxwell than Maxwell was over Kepler, shouldn't we expect more than the 75% improvement at 750 Ti's power envelope?

If not with Pascal, your dream should come true with Volta in 2017. Then you only need a 40% generational improvement to match 280X at 750 Ti's power requirement, since 1.4 * 1.4 = 1.96, giving you twice the performance over 750 Ti.

 

[The Sauce]

A little bit better than AMD's offering.

QFT

 

[moonbogg]

I have dreams and visions of a dual pascal (big die) SLI + Skylake-E 8 core rig with stupid fast DDR4...oh baby the only way that isn't going to be my rig is if I get hit by a bus or something.

 

[Arachnotronic]

I have dreams and visions of a dual pascal (big die) SLI + Skylake-E 8 core rig with stupid fast DDR4...oh baby the only way that isn't going to be my rig is if I get hit by a bus or something.

If I get hit by a bus, I got your back!

 

[guskline]

Nice dream moonbogg:thumbsup)

 

[Rvenger]

Dual GPU Pascal and not over $1200.00 and I would be onboard.

 

[poofyhairguy]

Well, you may not get 280X level performance but you could get pretty close. When 750 Ti was released, it was 75% faster than GTX 650 which still required a 6-pin connector. HD 7750 was a few percent slower than GTX 650 but didn't require that 6-pin (at the time, Techpowerup called it "[the] most power efficient graphics card in the world").

R9 280X is about 100% faster than GTX 750 Ti, and Pascal should offer a pretty good jump forward in efficiency thanks to the new process node. So... if Pascal will be a bigger jump over Maxwell than Maxwell was over Kepler, shouldn't we expect more than the 75% improvement at 750 Ti's power envelope?

If not with Pascal, your dream should come true with Volta in 2017. Then you only need a 40% generational improvement to match 280X at 750 Ti's power requirement, since 1.4 * 1.4 = 1.96, giving you twice the performance over 750 Ti.

2017? I will take it.

 

[zlejedi]

I think people will be disappointed by big Pascal due to super inflated expectations.

GM200 is 600mm^2 of pure gaming power - even with process shrink it won't be easy to offer some super huge gains when on fresh process and when it has to pack dual precision for professional workloads.

 

[steve wilson]

Yes. Guru3d has a rumor on their news page that the big Pascal GPU will be released in April this year. They always cost at least $1,000.00, but remember, the same cards used to cost about $500.00.

Thanks for the help... looks like I'll be waiting a while until I get to use my VR headset

 

[moonbogg]

Nice dream moonbogg:thumbsup)

I dream big like Luther.

 

[Flapdrol1337]

I think people will be disappointed by big Pascal due to super inflated expectations.

GM200 is 600mm^2 of pure gaming power - even with process shrink it won't be easy to offer some super huge gains when on fresh process and when it has to pack dual precision for professional workloads.

Probably, but with HBM they can save die space too.

 

[Railgun]

Better than Kepler. Not as good as Volta.

 

[DooKey]

And filthy. Don't forget filthy.

Don't forget "Take your stinking paws off of me you damn dirty ape!"

 

[desura]

As to why the power draw is important to you?

At the time I was running an underpowered PSU. I've since replaced it, but overall lower power = more reliability and less trouble.

 

[Good_fella]

I think people will be disappointed by big Pascal due to super inflated expectations.

GM200 is 600mm^2 of pure gaming power - even with process shrink it won't be easy to offer some super huge gains when on fresh process and when it has to pack dual precision for professional workloads.

You mean same as people expected that Fury X will be 40% faster than Titan X?

 

[Innokentij]

You mean same as people expected that Fury X will be 40% faster than Titan X?

People did?

 

[Pariah]

It used to be when you buy a high end GPU, you literally forgot about it for the next 2 years because you know nothing better was coming out for a long time.

There was pretty much nothing in your post that was factually accurate, however I will limit the response to the key point you attempted to make. Since the release of the original Voodoo graphics chipset in 1996, the first legitimate PC 3D accelerator, there has never been a single GPU flagship that made it 2 years without being surpassed. Both Nvidia and AMD have consistently released a new flagship every year, though not necessarily exactly 12 months after the previous release.

7900 GTX - March 2006
8800 Ultra - May 2007
9800 GTX - April 2008
GTX 285 - Jan 2009
GTX 480 - Mar 2010
GTX 580 - Nov 2010
GTX 680 - March 2012
Titan - Feb 2013
Titan Black - Feb 2014
Titan X - March 2015

I can go back further, but the point is made. Going back a decade, except for the 580, Nvidia has released a new flagship in the Feb-May range every year. Where are these mythical 2 year flagship cards that we were accustomed to having?

The way things are now, there is something a little faster coming out every 6 months or so and it doesn't matter if you buy high end or mid range. They release them in such a way so that next gens mid range comes out right after this gen's high end, so you buy the new mid range, but then next gen's high end comes out (which you thought you just bought a few months ago but you were mistaken because you really bought the mid range) and you wouldn't know the difference because the cost of mid range and high end cards are the same.

Having just shown you how predictably Nvidia releases flagships, if you were unaware that a Titan variant (followed by a 980Ti) was going to be released in the early part of 2015, you were part of a very exclusive group of enthusiasts with their head in the sand.

The rumors of a Pascal Titan being released in April probably have some basis in looking at Nvidia's past release schedule. I have serious doubts that Nvidia will have a Pascal Titan in April, but years of history hint that they could have something up their sleeves to trump the Titan X. Who knows now, but we all should know in a few months.

 

[moonbogg]

That was a big post. My point is still legit. How could you think you could negate my legit knowledge with your wall of text?
It used to be you buy the best an architecture has to offer, and you buy it once. There might be a slightly better product come later, like the occasional ultra variant or an increase in ram and clockspeed or something like that, but pretty much when you bought a high end card, you knew it was the big daddy and you could forget about it until the next legit card was released.
Now, they sell you a mid range first, then a high end, then a mid range, then a high end. Would you ever have bought like that before? I wouldn't have. Why would I buy a flagship card and then replace it with next gen's mid range card? You get less performance upgrade like that and you are compelled to buy the real flagship when it finally does get released. This means enthusiasts often buy twice instead of once like they used to.
This has been pointed out many times and is a common observation now ever since the 600 series screwed everything to hell. I don't know why you are pretending like this isn't the case. You seem to be interested in nit picking my post and pointing out little flaws when none of that crap takes away from the main point that was being made.

 

[MrTeal]

Well, you may not get 280X level performance but you could get pretty close. When 750 Ti was released, it was 75% faster than GTX 650 which still required a 6-pin connector. HD 7750 was a few percent slower than GTX 650 but didn't require that 6-pin (at the time, Techpowerup called it "[the] most power efficient graphics card in the world").

R9 280X is about 100% faster than GTX 750 Ti, and Pascal should offer a pretty good jump forward in efficiency thanks to the new process node. So... if Pascal will be a bigger jump over Maxwell than Maxwell was over Kepler, shouldn't we expect more than the 75% improvement at 750 Ti's power envelope?

If not with Pascal, your dream should come true with Volta in 2017. Then you only need a 40% generational improvement to match 280X at 750 Ti's power requirement, since 1.4 * 1.4 = 1.96, giving you twice the performance over 750 Ti.

This is all absolutely seat of the pants, but based on the demo RTG showed of Polaris playing Battlefront at the same settings and framerate as a 950 while consuming 86W at the wall vs 140W, it's likely P10 is at least twice as efficient as a 950 in certain situations. A stock 280X is about 30% faster than a stock 950, and a stock 950 pulls about 70% more power than a 750Ti.

I don't see a strong reason at this point why nVidia would have worse perf/watt than AMD in the next gen, so I wouldn't see it being big stretch for around 280X performance within a 75W envelope, especially in 2017 games.