Nvidia Predicts 570X GPU Performance Increase

[lejay]

[citation][nom]twat_fandude[/nom]LeJay: Are you an electrical engineer? [/citation]
No
[citation][nom]twat_fandude[/nom]
Are transistors perfectly square?
[/citation]
No, I don't think so.
[citation][nom]twat_fandude[/nom]
Do they leave proportionately less space between them when they are smaller?
[/citation]
I have no idea, I'm not an electrical engineer. (If that's what you have to be to know this stuff).
[citation][nom]twat_fandude[/nom]
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.
[/citation]
I tried that for around 30 secs and gave up. Realized it was irrelevant for the discussion. spanky deluxe did a rough calculation assuming perfect squares back to back. Spanky_McMonkey attacked this. Now, he did not point out that they are not perfect squares, nor did he point out the space between them were not proportional. He pointed out that when you decrease all sides of a square with 1/2 you get 1/2 the space. That is not correct, you get 1/4. I happened to find it slightly amusing that he wrote such an arrogant attack and made such an "epic fail" himself.
[citation][nom]twat_fandude[/nom]
Right now you're just some smug little twat who can do multiplication, I'm not particularly impressed.
[/citation]
I actually used a calculator, so I guess I'm just a smug little twat.

 

[Impulse Fire911]

i find it hard to use the picture shown on the article to claim as 570x times better. but idk soon it will be virtual reality so like you know beat that?

 

[Guest]

LeJay: shouldve_quit_after_8800, spanky_mcmonkey, twat_fandude, I_know_Im_right_thanks, and many, many other anonymous posters with cool screen-names and witty posts might all be the same person

"I tried that for around 30 secs and gave up."

Well, lucky for you I did it myself, and will repost it below. Please note how it lends credibility to my previous post.




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

 

[geoffs]

[citation][nom]righter_by_the_minute[/nom]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,...[/citation]Windsor is also a dual core and deneb is quad core, at least 2x the logic, not counting that it's a completely redesigned, more parallel core in deneb.

Try comparing chips that are at least close in architecture, like comparing Windsor to Brisbane:
Athlon X2 5000+ (Windsor), 90nm SOI, 154M trans, 2x128k L1, 2x512k L2, 220 sq mm.
http://www.techpowerup.com/cpudb/347/AMD_Athlon_64_X2_5000+.html
Athlon X2 4600+ (Brisbane), 65nm SOI, 154M trans, 2x128k L1, 2x512k L2, 126 sq mm.
220/126 = 1.74x

or compare Agena to Deneb, both are quad core, 512k L1, 2M L2.
Phenom X4 9550 (Agena), 65nm SOI, 2MB L3, 463M trans, 283 sq mm.
http://www.techpowerup.com/cpudb/496/AMD_Phenom_X4_9550.html
Phenom II X4 805 (Deneb), 45nm SOI, 4MB L3, 758M trans, 258 sq mm.
http://www.techpowerup.com/cpudb/610/AMD_Phenom_II_X4_805.html
463/283 = 1.636M trans per sq mm
758/258 = 2.938M trans per sq mm
2.938/1.636 = 1.796x as many transistors per sq mm.

1.75x-1.8x per generation is approximately 3.15x going from 90nm to 45nm. That's without any high-K metal gates, or other significant changes in the process technology, just 90nm SOI, to 65nm SOI, to 45nm SOI from one vendor.

1. XXnm refers to transistor width2. Transistors are not square, but much longer than they are wide3. Transistor length does not shrink proportionately to width[/citation]That's the only thing you've said about ICs that is accurate. However, gate length can and does shrink, just not proportionally. Other factors affect total size, that's why Moore's law has kept up for 40 years, and it's virtually certain that it will continue for at least another 6 years.

None of that even considers the possibilities of 3D chips or optical logic, which is probably at least 5 years out, more likely 10+.

 

[carlhenry]

nvidia will produce green lightsabers and amd the dark red light sabers. we will see star wars after 6 years. yay! may the force be with you. which side are you on?

 

[Guest]

Geoffs: Optical logic has never proven itself possible, yes they can channel "nano-light", but until something akin to the transistor is developed **AND** can support a fast switching speed, then optical logic is a pipe-dream at best. Even if 3d chips are made, there are still inherent limitations there, even if it does bring improvements and shorten the distance that electricity has to flow, each transistor still has a more-or-less fixed electrical draw, meaning that you can only have so many of them firing at one time and still stay within a reasonable power envelope.

 

[geoffs]

[citation][nom]still_not_6x_is_it[/nom]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 55millionPentium D 280mm2 376million145/280 = 0.5178571430.517857143 * 376million = 194.714285714million194.714285714 / 55 = 3.54[/citation]Don't know where you got your die size for The Pentium D 9xx, but apparently, you can't be trusted to do it correctly. It's 162 sq mm, not 280.

145/162= 0.895
.895 * 376M = 336M
336M/55M = 6.1x going from 130nm to 65nm.

Or compare the Northwood to it's immediate successor, the 90nm Prescott.
From http://techreport.com/articles.x/6213/1
"Not only is the 90nm process smaller, but Intel is also manufacturing Prescott using seven layers of copper interconnects, instead of the six used at 130nm. All told, the changes shrink the Pentium 4's die size to 122 mm2, from 145 mm2 for Northwood—this despite the fact Prescott's transistor count is 125 million, over twice Northwood's 55 million transistors."

Do all the math and you get 2.7x increase in a single generation shrink from 130nm to 90nm.

Also, look at the Intel generated slide included here
http://www.overclockers.com.au/image.php?pic=articles/452447/mooreslaw.jpg
that shows the gate length shrinking proportionally to the feature width.

 

[lejay]

[citation][nom]anonymous_tranny_thing[/nom]LeJay: shouldve_quit_after_8800, spanky_mcmonkey, twat_fandude, I_know_Im_right_thanks, and many, many other anonymous posters with cool screen-names and witty posts might all be the same person "I tried that for around 30 secs and gave up."Well, lucky for you I did it myself, and will repost it below. Please note how it lends credibility to my previous post.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 = 2.647813188[/citation]
You don't address a single point in my post, compare two very different cpus and still prove my point. If cpus (weren't we talking gpus?) were 1 dimensional it would be impossible to get more than 2x increase. So it is not 1 D... What is the next possible dimension? And how does 2D squares behave? Thank you. We were discussing math, so stop pointing to random cpus.
If you want to address someone who is talking about the actual increase in transistors pr. nm^2 when all practical issues are considered leave me out of it. I was talking math and made that very clear.

 

[Guest]

I made an account just to tell a couple of you that if you can do math yourself dont criticize someone else's math.

 

[lejay]

[citation][nom]mertzj_90[/nom]I made an account just to tell a couple of you that if you can do math yourself dont criticize someone else's math.[/citation]
Because everyone is entitled to their own opinion on what is mathematically correct?
Interesting, maybe some right wing Christians will start teaching that pi=3 .

 

[geoffs]

[citation][nom]2many_screennames[/nom]Geoffs: Optical logic has never proven itself possible, yes they can channel "nano-light", but until something akin to the transistor is developed **AND** can support a fast switching speed, then optical logic is a pipe-dream at best. Even if 3d chips are made, there are still inherent limitations there, even if it does bring improvements and shorten the distance that electricity has to flow, each transistor still has a more-or-less fixed electrical draw, meaning that you can only have so many of them firing at one time and still stay within a reasonable power envelope.[/citation]You mean this optical transistor?
http://www.sciencedaily.com/releases/2009/07/090702080119.htm

Yes, it's a research project and it requires cooling to almost absolute zero. It might never make it to commercial use at room temperature, but the basic components exist today.

That's why I said the "possibility of", there is no certainty that those technologies will ever be viable, only that they show potential.

Yes, electronic transistors draw power, it's called leakage current. There are ways to minimize that, including high-K dielectrics, better insulators, lower capacitance circuits, powering down inactive circuits, asychronous (clockless) logic, increased parallelism with reduced clock rates, etc. There is also "reversable logic". You should read up on those.

 

[wolfseeker2828]

Wasn't this also the guy who said that the CPU's job was soon going to be offloaded to GPUs due to NVidia's CUDA? Well I took out my CPU from my computer, because he is always right, and it didn't work. So I'll believe him when my computer can run Crysis without my processor.

 

[idisarmu]

[citation][nom]snotling[/nom]probably 570x compared to an intel i945G[/citation]

Nah... that already exists: ATi and Nvidia both have chips like at already: The HD3450 and the 9400GT respectively.

AMAZING!

 

[geoffs]

[citation][nom]wolfseeker2828[/nom]So I'll believe him when my computer can run Crysis without my processor.[/citation]You have a computer that can run Crysis? You must be rich.

 

[mysticalzero]

"that GPU computing will dramatically increase over the next six years, a mere 570 times that of today's capabilities in fact, while CPU performance will only increase a staggering 3x in the same timeframe."

If that's the case, then wouldn't the GPU be heavily bottlenecked by the CPU?

 

[marraco]

Historically, nvidia GPU had improved 2X -on average- each year, so in 6 years, nVidia can improve 2^6=64X.

But because of poor competence from ATI, NVidia did not released all of his available technology last year, (like use of DDR5) so it conserves an extra year under de hood, ready to be released.
It can make 2^7=128X improvements in 6 years.
let's suposse that nVidia is counting in double chips video cards. then it can reach
2x2^7=256X
The only way to break 500X would be use of SLI cards... 2 x 2 x 2^7=512X

maybe possible, but not exactly what nVidia infers from what i read this article.

And even if possible, we know that NVidia will not release the technology if ATI/Intel does not challenge them.

By the way, the 22nm quantum noise barrier is not the end of the line. In 6 years, multigate transistors may be ready for production.
http://en.wikipedia.org/wiki/Multigate_device

also, nVidia can count not only on transistor shrinking, also can reach the 4 Ghz, and DDRx logic (Not justDDR memory, but also logic)

 

[lejay]

[citation][nom]mysticalzero[/nom]"that GPU computing will dramatically increase over the next six years, a mere 570 times that of today's capabilities in fact, while CPU performance will only increase a staggering 3x in the same timeframe."If that's the case, then wouldn't the GPU be heavily bottlenecked by the CPU?[/citation]

Depends on the task. I think cpu usage is pretty low in something like video encoding.

 

[mikenygmail]

This is obviously nonsense, and high end GPU performance will not increase 570X from current levels.

 

[mikenygmail]

Definately not within 6 years anyway!

 

[Guest]

ya that's awesome but I wish that they would get there drivers more stable. I have been using Win 7 evaluation copy build 7100 and can never get it to use the newest drivers because I get the BSOD all the time I can only use Driver 185.85 to work with my computer without any problems.

 

[FallenSniper]

I wish it was 570x performance by next year!

 

[GullLars]

I suspect 570x reffers to FLOPS.
The HD 3xxx series had a max of 497 GigaFLOPS, one year later 4xxx had 1,2 TerraFLOPS, and this fall 5xxx is expected at 2+ TFLOPS. If this continues you'd expect to se roughly 50x more FLOPS in 6 years. With refined arcitecture, higher clock speed, and more memory bandwith you could perhaps add in a small multiplier to that, but i doubt 11,5x = 570x.

 

[lejay]

The difference between 3xxx and 4xxx is because of "refined arcitecture, higher clock speed". You cannot count that twice.

 

[GullLars]

If you refine the arcitecture to do more work pr clock, that is one optimalisation. If you then in addition turn up the clock (core, shader, or memory), you get even more. If memory bandwidth then is the bottleneck in some cases, you can get even more by increasing the bus width to the memory.

 

[lejay]

The square root of 16 is 4.