The GT300/Fermi Thread - Part 2!

[saaya]

Seeing how vast majority of programs, and even games are barely even optimized for 2 cores, you can imagine how much ingenuity this requires on GPU side.

the tasks that are performed on the cpu and gpu are completely different though, the cpus strength is serial work, the gpus strength is parallel work... thats why coding games or apps to use several cpu cores is just as much as a pain in the lower rear side as it is to code apps that use the gpu to perform serial work, which, most apps need at some point... thats why gpus get their own cache and get their task schedulers reshuffled and upgraded and will probably even get branch prediction, while cpus get avx and beefed up double pumped execution cores ala gpu... thats at least how i understood it...

show me a framerate comparison at the dragon now

oh god, i cant believe this low blow marketing actually works... you really care how cards perform in some synthetic highly unrealistic scenario? this reminds me a lot of memory benchmarks where one set is 2x as fast as another but you wont be able to find a real world app that shows a perf gain past .x%...

GTX 480 has 480sp, as charlie said,haha

mhhhh that increases chances for people to actually be able to find one, provided they have enough money and fast index finger reflexes to hit the BUY button fast enough when one shows up in the shops...

but it also means lower performance...
im sure that nvidia didnt like this... either yields were too bad or maybe they figured out that 480sps (8%) allow them to clock the card more than 8% higher (~50mhz) so perf wise it makes more sense... or maybe mhz gives more of a perf boost than more sps... that would make sense since 500sps is quite some parallel power that isnt easy to feed i assume...

http://www.youtube.com/watch?v=gNIPc...eature=related

Came across this and LOL'd

hahahah, that is hilarious
i wish more pr campaigns would be based on funny stuff like this... it doesnt matter what product you own or think is better, those clips are simply hilarious

Right. Even if there is only one working 512SP-Part on the Wafer, it doesn't make sense not to sell it.

thats provided it not only has 512 working sps but 512sps capable of running at decent clockspeeds AND being able to do so without triggering a thermo nuclear meltdown... jokes aside, it actually makes little sense to sell those parts as vgas... youd think that anything above 448sps would be sold as tesla cards, as they retail for 2500us$+... the only reason to sell those big fat gpus in retail as vgas is for pr purposes... which is confirmed by their, relatively, low volume...

AFAIK we don't even have a confirmed 480 PCB picture.

yeah we do... you didnt see it?

I wonder how many chips they have that actually work like that though, obviously only a couple but i wonder exactly how many

i think there are quite some, but how many can work at those specs at room temps and with acceptable cooling is probably very lower... 750mhz plus 512cores... im sure they have at least a hand full of those since they ordered 10k wafers, right? so even with 0.1% yields they would have 10 of those cards now...

Hmm. 3? One for Jensen, one for Igor. Jensen's is SLI.

pfff, igor gets a single gts240 card with 480 heatsink glued to it... as if hed even notice the difference...

so wait, the highest launch part is the 480, and it wont have 512sps...
so there wont even be GTX480 Ultra reviewer editionTM cards?

hmmmm i wonder if some people will actually buy a tesla card for 2500$ to get the additional 16sps and win in some benchmarks, wouldnt surprise me

about gtx480 temperatures... did you guys see the fermi approved thermaltake case at cebit? it has a huge duct for the vgas in it... its really cool if you ask me, literally!
if id use a case id probably use that one!
but it sounds as if this duct is required to run 3 480s next to each other without any plastic melting... thats less cool... literally...

http://www.semiaccurate.com/2010/03/...ase-dissected/

[mrcape]

http://www.youtube.com/watch?v=gnipc...eature=related

came across this and lol'd

that was hilarious! Not taking any sides though, both cards are cool with me.

[Behemot]

but it sounds as if this duct is required to run 3 480s next to each other without any plastic melting... thats less cool... literally... [/url]

This...

I assume the temps won't be so bad at all, mainly due to better cooler desing and the fan will probably spin more than 3000 RPM. But what actually is problem is Tri-SLI IMHO the holes in PCB are made for Dual-SLI, but they are useless for Tri-SLI, the middle card will vaporize

[LiquidReactor]

Says who? You who said Fermi is G80 size?

AFAIK we don't even have a confirmed 480 PCB picture. Then again, you have magical vision that can see the die size under the IHS

The pcb was posted a few pages back several times

Oh and I said Fermi was close to G80 size. Judging by IHS it has to be at least smaller then GTX 285 and since you have no better counter then IT HAS TO BE BIGGER, CHARLIE SAID SO I'm just going to go by IHS size.

[Olivon]

http://www.youtube.com/watch?v=gNIPc...eature=related

Came across this and LOL'd

Really fun as the Hitler's video about Fermi.

[zalbard]

yeah we do... you didnt see it?

I think his point is that it has not been confirmed to be a real final PCB. More of a 'leak' type information.

so wait, the highest launch part is the 480, and it wont have 512sps...
so there wont even be GTX480 Ultra reviewer editionTM cards

I think they may make an Ultra card at some point in time, once yields are decent. I really don't see why not if all current chips are being made with 512sps and some of them are getting disabled...

about gtx480 temperatures... did you guys see the fermi approved thermaltake case at cebit? it has a huge duct for the vgas in it... its really cool if you ask me, literally!
if id use a case id probably use that one!
but it sounds as if this duct is required to run 3 480s next to each other without any plastic melting... thats less cool... literally...

http://www.semiaccurate.com/2010/03/...ase-dissected/

I do not think this is a requirement, though. Could be an indicator, of course... But probably is just a marketing trick. And a cool feature!

[zerazax]

The pcb was posted a few pages back several times

I must have missed where it was a confirmed 480 one, and not just a 470 one with blurred info

Oh and I said Fermi was close to G80 size. Judging by IHS it has to be at least smaller then GTX 285 and since you have no better counter then IT HAS TO BE BIGGER, CHARLIE SAID SO I'm just going to go by IHS size.

IHS has nothing to do with die size, besides provide a maximum boundary.

IHS is related to bus width due to pins. 384-bit vs. 512-bit is the big reason why a GT200 and Fermi is different in size, and why Fermi's IHS appears close to a G80, which was also 384-bit.

But again, IHS size is not proportional to die size. But keep ignoring that fact! (And what does Charlie have to do with any of this? nice random tangent )

[SKYMTL]

but it sounds as if this duct is required to run 3 480s next to each other without any plastic melting... thats less cool... literally...

http://www.semiaccurate.com/2010/03/...ase-dissected/

Not true. Thermaltake's reps have been talking about this since before the GF100 series was announced. It (or a modification thereof) will be installed on other cases as well, not only NVIDIA editions.

All you are seeing is an evolution of the "ATC" design used on the Spedo Advanced.

You're reading far too much into a simple article.

[saaya]

I think his point is that it has not been confirmed to be a real final PCB. More of a 'leak' type information.

I do not think this is a requirement, though. Could be an indicator, of course... But probably is just a marketing trick. And a cool feature!

requirement, def not... but i wouldnt be surprised if nvidia recommends special cases for sli or tri sli... then again, im surprised ati didnt do this for the 4870x2 and 5970 since they dont fit in quite some cases and run pretty hot as well...

Not true. Thermaltake's reps have been talking about this since before the GF100 series was announced. It (or a modification thereof) will be installed on other cases as well, not only NVIDIA editions.

then why isnt there an ati version of it? why is the case only fermi approved/certified/recommended whatever?

its not much, but its a hint at 480s running pretty hot... not that we wouldnt know this already...

It appears that NVIDIA has announced a thermal specification for their upcoming GF100 gaming graphics cards and this is one of the cases that meet those requirements.

http://forums.legitreviews.com/about25148.html

[SKYMTL]

requirement, def not... but i wouldnt be surprised if nvidia recommends special cases for sli or tri sli... then again, im surprised ati didnt do this for the 4870x2 and 5970 since they dont fit in quite some cases and run pretty hot as well...

then why isnt there an ati version of it? why is the case only fermi approved/certified/recommended whatever?


http://forums.legitreviews.com/about25148.html

Thermaltake marketing at its finest. All that's missing is the poorly translated press release

Cooler Master had their NVIDIA Edition of the Stacker and and ATI Edition of the HAF. I am sure Thermaltake will have an ATI Edition of something as well.

There is no doubt that the cards will be hot-running judging from the heatsink design. However, most cards from what I know have certain guidelines when it comes to designing "approved" / "certified" enclosures. Take the HD 5970 and older GTX 280 for some examples. Usually those specifications simply include the provision for front and back exhaust fan in order to get proper air circulation within the enclosure. Maybe the certification criterea is different for the GF100 but I can't see it changing that much from what it already is for the likes of the GTX 295 and other high-TDP cards.

However, to assume a case will "need" all these add-ons in order to accomodate the heat produced by three GTX 400 series cards is a bit of a stretch IMO.

[saaya]

hmmmm i guess its just thermaltake marketing then that is trying to make their case look better by hinting that you NEED a special case for fermi... im surprised nvidia cooperates with them without making sure they dont spread nonsense like this that makes their cards look bad, nvidia is usually very fast with stuff like this...

[-Sweeper_]

oh god, i cant believe this low blow marketing actually works... you really care how cards perform in some synthetic highly unrealistic scenario? this reminds me a lot of memory benchmarks where one set is 2x as fast as another but you wont be able to find a real world app that shows a perf gain past .x%...

not that I care, but he was comparing tessellation performance..

[mindfury]

Rumor:GeForce GTX 470 will sell for 2499¥(366$) in China.

http://diy.yesky.com/vga/259/11163259.shtml

[Jamesrt2004]

Rumor:GeForce GTX 470 will sell for 2499¥(366$) in China.

http://diy.yesky.com/vga/259/11163259.shtml

don not like rumour part!

[saaya]

Rumor:GeForce GTX 470 will sell for 2499¥(366$) in China.

http://diy.yesky.com/vga/259/11163259.shtml

no way jose...
nvidia doesnt underbid ati... ever... even when they are notably slower... and in this case they seem to be close to 5870 performance... plus they are supply limited... not gonna happen...

[annihilat0r]

470 will have a MSRP of between $300 and 400. It won't be both slower and more expensive than an ATI offering. Nvidia might have failed to adjust some prices after an ATI move but this is a new card and you can be sure it's going to be priced accordingly.

Of course, with low availability there'll be inevitable price gauging but the original price will be between $300 and $400.

[saaya]

but this is a new card and you can be sure it's going to be priced accordingly.

well thats contradicting... new cards always cost more than the competition even if they are slower... theres always a 10%+ tax for every piece of hardware that is new...

Of course, with low availability there'll be inevitable price gauging but the original price will be between $300 and $400.

well, 399$ is between 300 and 400, so yes, what your saying is not entirely untrue
but just think about it... there will be enough people willing to spend 400$ for it and nvidia has very few cards... why should they set the msrp that much below the expected demand based price? their aibs would barely make any money on the cards, especially considering the low volume, and the etailers and retailers would make a fortune on the cards by selling them for msrp+50-100$... why would nvidia do this? nvidia sales guys are brilliant! just look at how well they have done in the past! they wont do something like this...

the 470 will only go for 350$ if the 5870 drops to 350$ as well or even lower... but i dont think that will happen... it would be ati shooting themselves in the foot, cutting their wins on millions of cards for some few thousand cards nvidia drip feeds the market...

[SKYMTL]

hmmmm i guess its just thermaltake marketing then that is trying to make their case look better by hinting that you NEED a special case for fermi... im surprised nvidia cooperates with them without making sure they dont spread nonsense like this that makes their cards look bad, nvidia is usually very fast with stuff like this...

I don't think it makes things look bad unless you blow it out of proportion by assuming ducts, oversized fans and whatnot are NEEDED for cooling the GTX 400 series.

Personally, I like the Thermaltake addition simply because it will help cards that don't have the ability to exhaust hot air directly outside the case.

[annihilat0r]

well thats contradicting... new cards always cost more than the competition even if they are slower... theres always a 10%+ tax for every piece of hardware that is new...

nah, I meant that Nvidia will price a card fair "initially". After an ATI move (new card or price cuts) Nvidia is known for not adjusting prices all the time, but when pricing a new card, they'll make it fair. Or close to fair

I agree that GTX 470 might be $399. But that's not really far off from the $360 figure which you seemingly found impossible.

[saaya]

Nvidia is known for not adjusting prices all the time, but when pricing a new card, they'll make it fair.

nvidia offering fair prices? cant remember a single instance.. seriously :P
nvidia was always about 5% faster, better (according to marketing) and 20% more expensive...

I agree that GTX 470 might be $399. But that's not really far off from the $360 figure which you seemingly found impossible.

i find it hard to believe that 470s will go for less than a 5870... especially with nvidia trying to convince us so hard that its better than a 5870 and rumors of a very limited supply of cards...

[***Deimos***]

The proportion is also affected by the speed of the CPU. i.e the faster the CPU, the smaller the serial part becomes. Yes, it's just a fancy way of saying that as you remove the CPU bottleneck the proportion of parallelizable work tends toward 100% of the total. But even on the GPU a lot of stuff is serialized - geometry setup for example. Hence the multiple setup engines in Fermi. The more stuff you parallelize, the easier it is to scale performance.

Well good at least nVidia is working on parallelizing more of the pipeline.
Serial would be better, but clocks have been stuck in 600-700 range for half-a-decade, so like CPU, only way around it to ensure GPU is not starved, is parallel setup engines.. and all the coherency issues assossiated with it.

not much, gpu applications like graphics are embarassingly parallel and this parallelism increases with problem size. i.e. if you double the pixel count you double parallelism. this law sounds very grim but truthfully its not. gpu's are already running thousands of threads to hide latency.

How easy is to to scale from thousands of threads to millions? Buffers and register files are already enormous.

Problem is that graphics is assumed to be infinitely parallel. But you can only work on about 2 million pixels at a time, before starting next frame. What happens with 20 million triangle tessellation demo running at 800x600... 40 triangles/pixel on avg - which you cant distinguish. Of course we're still some years away.

But perhaps the greatest challenge is the efficiency ceiling. Some shader applied to block of 15 of 16 pixels is ~94% resource efficient. But extrapolate to dozens of shaders, and 0.94^12 doesn't look so good. GPUs already use an enormous amount of optimizations, from non-sqrt Z calc, to buffer compressions, to normal/bump/stencil maps, mip maps and LOD. It seems only natural that its more and more difficult to improve this fine tuned model.

[Olivon]

From the same guy PHK on bbs.expreview.com forum.

Don't know what card is, maybe GTX 470 but not sure

[ajaidev]

$366 for a GTX 470 may be correct.

You can already get a 5870 for $380

A GTX 470 cant be more expensive than the 5870 because it does not seem to perform much better than 5870 in normal cases.
I expect GTX 470 to cost around 5870 at most $10-$20 less...

[Blacky]

When does "officially" the NDA ends? Fermi still clueless, I still think nvidia is not going to make it for a march launch (maybe a paper launch), plus I still think performance is going to disappoint for the price they will ask, not mentioning the power hog and hot the card will be, thats just me I could be wrong though (performance wise).

[Dimitriman]

From the same guy PHK on bbs.expreview.com forum.

Don't know what card is, maybe GTX 470 but not sure

Considering its another 2500 resolution opportunistic comparison I'd be pretty sure that's a 470.