Nvidia Predicts 570X GPU Performance Increase

[buzznut]

Why would they bother? All they need to do is keep rebranding old tech over and over again and people (suckers) will continue to buy it. This certainly doesn't look like what nvidia has been doing for the past two years.

My guess? In two years we'll still be buying some version of the 8800GT. In five years they;ll probably still have some version of the 280 gtx kicking around.
Of course it won't be called that and it will be in a nice new shiny package, but they'll still be trying sell that old tech.

 

[raceway99]

this bodes well for the gt300 series, cant wait!

 

[crom]

So basically Nvidia is telling us to wait before buying a GPU because they're going to go up exponentially in a short period of time. Sounds like a bad business announcement to me.

 

[fulle]

The only way I see 570x possible... is if they have something in the works to produce Graphene based chips. There's tech for Graphene to do some kind of frequency doubling, and the transistors are able to get down ridiculously small. A sheet of Graphene can be just 1 atom by 8 atoms.

Now, in CPUs, IBM has gotten the frequency up to 26GHz, and people are talking 1000GHz chips. So... with a tech that can give at least 10x the frequecy, and give such tiny transistors. Seems easily possible.

With 22nm silicon chips, I'd call bullshit... but, I have my fingers crossed that Graphene chips in 2015 might be a real possibility. Its hard to find any real information on how things are progressing.

 

[traviso]

[citation]16nm die fab = 16x as many processing units + faster processing units[/citation]
Where did you make up this horribly wrong math? 16nm refers to size, today's GPUs are 55nm, so a move from that to 16nm makes the math roughly:
55 / 16 = 3.5x more room in the same die

So ignoring fact that hte GPU gets more efficient, you only gain less than 4x the speed improvement (or 400% but notice the quote is 570x not 570% which I believe is a mistake, he meant 570%)

Generally speaking, every new generation only offers a 50% boost from the previous generation, or 1.5x faster

 

[Guest]

Are we even going to have dedicated GPUs and CPUs in six years? There's been all this talk of the two merging together. In that case the performance increase would be more like (570 + 3) / 2 = 286.5x.

 

[bounty]

Tom's writers need to take some comprehension classes or stop intentionally starting flame wars.

"Huang showed a number of real-world examples of heterogeneous computing. The range of applications that are accelerated by GPU computing extend from oil and gas exploration, to interactive ray tracing, to the simulation of realistic looking directed fire for movie magic.

Looking six years into the future, Huang believes that GPU Compute can offer 570 times the capability we have today, while pure CPU performance growth might only offer three-times the performance"

He's talking about using Tesla/Cuda for non-UT2004 applications @#$#@^^%^. Crazy scaling has already happend for some apps, and is getting better. http://www.elcomsoft.com/edpr.html In 6 years I can totally see 570x say, Q6600 performance in that app. Sure it's market speech, and you fudge a little this way and that, but basically we will see a massive scaling for certain tasks that use the GPU in new ways.

 

[bounty]

Office 2007 password "recovery" is already 50x faster with tesla over a Q6600. Now imagine 6x performance in 6 years for Nvidia (easy). So you can kinda say we're at 300x performance. Multiply that by 2x for more effecient software (this is new ground so impovements here are possible) now we're talking 600x performance.

(p.s. Intel has roadmaps out to 4 nm. Those crazy guys may have some secrets.)

 

[geoffs]

[citation][nom]snotling[/nom]probably 570x compared to an intel i945G[/citation]I think we already have that, it's called a Radeon HD 4870 or a GeForce GTX 285.

 

[geoffs]

[citation][nom]twat_fandude[/nom]LeJay: Are you an electrical engineer? Are transistors perfectly square? Do they leave proportionately less space between them when they are smaller? Go dig up any of AMDs or Intels 90nm and 45nm chips, and compare the number of transistors to the die size, then report your findings back here, including the math. Right now you're just some smug little twat who can do multiplication, I'm not particularly impressed.[/citation]LeJay is correct. Go look up the numbers yourself. If I'm not mistaken, this is tomsHARDWARE.com, where things such as feature size (process technology) and die size are discussed pretty regularly.

 

[geoffs]

[citation][nom]traviso[/nom][citation]16nm die fab = 16x as many processing units + faster processing units[/citation]Where did you make up this horribly wrong math? 16nm refers to size, today's GPUs are 55nm, so a move from that to 16nm makes the math roughly:55 / 16 = 3.5x more room in the same dieSo ignoring fact that hte GPU gets more efficient, you only gain less than 4x the speed improvement (or 400% but notice the quote is 570x not 570% which I believe is a mistake, he meant 570%)Generally speaking, every new generation only offers a 50% boost from the previous generation, or 1.5x faster[/citation]He didn't make it up, it's the way it works. 55nm or 16nm is the 'feature size'. ICs have basically 2 dimensions, length and width. Cut the feature size in half, you quadruple the number of transistors per unit of area, because you gain 2x in each dimension. 16nm vxs 55nm ICs can have (55/16)^2 = 11.8x as many transistors on the same size die, or they can have the same number of transistors on a die that is 1/11.8 the area.

The power will not change by 11.8x because you'll use lower voltages at smaller feature sizes, but it would take a significant breakthrough to put 11.8x the transistors in a 16nm IC without the power increasing by at least a factor of 4.

 

[geoffs]

Somehow, my response didn't get separated from what I was responding to. Here's the portion that is my response:

He didn't make it up, it's the way it works. 55nm or 16nm is the 'feature size'. ICs have basically 2 dimensions, length and width. Cut the feature size in half, you quadruple the number of transistors per unit of area, because you gain 2x in each dimension. 16nm vxs 55nm ICs can have (55/16)^2 = 11.8x as many transistors on the same size die, or they can have the same number of transistors on a die that is 1/11.8 the area.

The power will not change by 11.8x because you'll use lower voltages at smaller feature sizes, but it would take a significant breakthrough to put 11.8x the transistors in a 16nm IC without the power increasing by at least a factor of 4.

 

[jeffkro]

Its sure a good thing that I'm skipping this generation of graphics cards then. I figure by holding off I will get a Graphics card thats 250 times better.

 

[Corporate_goon]

Weren't we supposed to hit a quantum effects wall at 130nm, too? Shrinking below 22nm might be hard, but I doubt it's impossible.

Comment about this thread: It's always hilarious when someone starts ranting about someone else's grievous errors, when they themselves are hilariously wrong. Spanky_McMonkey, as has already been pointed out, we're talking about two dimensions, not just one, so shrinking to 22nm from 55nm will give you over 6x the space to fit cores, not 2.5x.

 

[Guest]

normal performance increase rate for computing commodities is around 2x per 18 months, which leads to about 600% increase over 10 years. So he is basically saying that GPU being a new playground, there might be a little bit of free lunch and we could see a slightly higher than normal performance growth rate. Big announcement, but actually pretty obvious for anybody in the field.

 

[Guest]

OK noobs, specifically Corporate Goon, Geoffs, and anyone else who didn't believe me:

A comparison of 90nm Windsor and 45nm Deneb:

Windsor 219mm2 243million transistors
Deneb 258mm2 758million transistors

//normalizing the die size difference
219/258 = 0.848837209
0.848837209 * 758million = 643.2million

//computing the difference in density
643.2million / 243million = 2.647813188


Much closer to 2x, isn't it? Apology accepted.


PS: All the same, your strawman argument doesn't begin to explain Mr. NvidiaCEO's claim...

 

[Guest]

Oops: I also meant to give a shout out to LeJay in the opening finger-pointing.


PS: Only a little bitch would vote that comment down.

 

[Studly007]

I think all you guys would be better off following Hynix R&D, press release statements, and scheduled projection. That would be a FAR more Accurate Indicator.

NVidia just puts Hynix parts together, whether it's one or two GPU's on one card, or fusing two cards together (GTX 295).

Now - the biggest wild card, that I think is an idea with the likes of Bill-Gates-payoff-potential - is the sole proprietor who finally cracks open parallel OS software programming, parallel video game rendering engines (imagine Crytek, LucasArts, UbiSoft, Infinity Ward, and Epic putting all their heads / smarts together), parallel software tools, and implements that into 64 bit programming language for current core i7 CPU's to utilize all their cores, and breakdown the rendering between the CPU and GPU setup to (I dunno, arbitrarily) 33% to 67%.

THAT WOULD KICK MAJORbooty. Can you imagine a core i7 (with all the irrelevant background OS apps turned off) ACTUALLY firing on ALL FOUR cores at 3 GHz tackling a third of the ACTUAL rendering work of a Quad SLI GTX 295 setup?

DUDE !!!! ........ SWEEEEEET !!!!
(Where's my Car?)

 

[geoffs]

[citation][nom]I_know_im_right_thanks[/nom]OK noobs, specifically Corporate Goon, Geoffs, and anyone else who didn't believe me:A comparison of 90nm Windsor and 45nm Deneb:Windsor 219mm2 243million transistorsDeneb 258mm2 758million transistors//normalizing the die size difference219/258 = 0.8488372090.848837209 * 758million = 643.2million//computing the difference in density643.2million / 243million = 2.647813188Much closer to 2x, isn't it? Apology accepted.[/citation]No apology was offered! I tried to be polite about this, but since you want to get nasty....

You !@#$%#$% moron, those two CPUs are different architectures with different amounts of L1/L2 cache and other significant differences. Not all transistors are the same size, nor do they scale exactly the same way. Those used for logic circuits vs those used for cache memory vs those used for power and clock distribution are all different sizes.

For an clear example of that, look at these two Intel Core 2 Duo CPUs.
65nm E4400, 167M transistors, 111 sq mm die
http://www.techpowerup.com/cpudb/519/Intel_Core_2_Duo_E4400.html
45nm E8600, 410M transistors, 107 sq mm die
http://www.techpowerup.com/cpudb/516/Intel_Core_2_Duo_E8600.html
More than 2x as many transistors on a slightly smaller die, and that's only going from 65nm to 45nm. You can't compare two different architectures on die size and transistor count alone! Find a chip that was scaled from one feature size to the next smaller WITHOUT changing it's number of processing units, cores, cache sizes, or adding or removing other features, and you'll see that for a 30% reduction in "feature size", you get about a 50% reduction in area.

PS: All the same, your strawman argument doesn't begin to explain Mr. NvidiaCEO's claim...

I wasn't trying to explain his 570x comment, just verify that the calculations "spanky deluxe" used are in fact correct. I never made any claims about Jen-Hsun Huang's 570x being plausible or not. Apparently, you not only don't know anything about microelectronics, but your reading comprehension fails you.

 

[Guest]

Since the voting down has begun...

Some of that extra .647813188 in the 2.647813188 density is due to the difference in cache size, cache is more transistor-dense(due to more uniform packing) than the rest of the CPU, so if you factor in Deneb's huge cache vs. Windsors more modest cache, it would be somewhat less than 2.6, I'm going to estimate about 2.3x, so, going forward, we can say:

1. XXnm refers to transistor width
2. Transistors are not square, but much longer than they are wide
3. Transistor length does not shrink proportionately to width
4. There will always be self-righteous people with "a little bit of knowledge"
5. A little bit of knowledge is a dangerous thing
6. Sore losers vote down posts that pwn their "little bit of knowledge"

 

[Guest]

Geoffs: Those are 2 different kinds transistors, High-K metal gate transistors are much shorter than what Intel was using at 65nm. Both AMD chips use SOI, the same transistors, if anything, the different architectures only serve to reinforce my point in the previous post. Probably everybody will switch to High-K metal gates, but that is not a common occurence, and those are probably going to be the shortest(relative to width) transistors to ever be made. At no other evolutionary step will you find such a differnce, I challenge you to compare Intel 130nm to Intel 65nm for comparison.

I don't understand what point you are trying to make there, getting a little flustered? Will 22nm chips use the same architecture as 45nm?

 

[descendency]

The real problem isn't 22nm or 16nm or anything else. It's 9nm. At that point, the laws of physics start to say it's impossible to shrink any further.

 

[geoffs]

[citation][nom]descendency[/nom]The real problem isn't 22nm or 16nm or anything else. It's 9nm. At that point, the laws of physics start to say it's impossible to shrink any further.[/citation]"laws of physics as we know them now". Could have made the same statement 10, 20, and 30 years ago about what looked like limits then.

Granted, the limits we're approaching now are based upon the size of atoms, and that's a pretty significant limit, but who knows....

 

[Guest]

Since none of you can apparently be trusted to do it, here is a comparison of a 130nm Northwood Pentium IV, and a 65nm Pentium D 900.

Northwood 145mm2 55million
Pentium D 280mm2 376million

145/280 = 0.517857143

0.517857143 * 376million = 194.714285714million

194.714285714 / 55 = 3.54

This strays pretty far from the previous 2x, however, the difference in cache is quite significant here, and this IS essentially the same architecture on the same transistors, although they had streamlined it and increased it's density from 130nm to 65nm.

 

[rooket]

will it be compatable with my 486