PDA

View Full Version : NVIDIA GF100 Previews


Pages : 1 2 3 4 5 6 7 8 9 10 11 12 [13] 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52

CaptNKILL
03-02-10, 10:38 PM
Gigabyte boxes.

http://vr-zone.com/forums/570761/gigabyte-gtx-480-and-gtx-470-retail-boxes.html

Interesting... the 470 and 480 both have 384bit buses and 1.5Gb memory. That's unusual for Nvidia.

Plus, the rumor mill has pointed toward the 470 being cut down like all of their previous cards:

As of right now the nvidia GPU comparison chart (http://en.wikipedia.org/wiki/Comparison_of_NVIDIA_Graphics_Processing_Units#GeF orce_400_Series) on wikipedia lists the specs as:

1280Mb memory, 320bit bus, 448 SPs, 56TMUs and 44 ROPs... but those are clearly wrong if the Gigabyte 470 box says that it has 1536Mb of memory and a 384bit bus.

This is all very interesting. The 470 must just be a lower clocked card with less SPs and the same number of ROPs... which means it could be a killer card if priced well.

EDIT: wait a minute, those boxes have typos! :rofl

The front of the 470 box says 1280Mb memory while the bottom says 1536Mb. :lol:

I guess the rumored specs are most likely true. How the hell could they let something like that slip? :p

shadow001
03-02-10, 10:41 PM
what are c++ hooks? :), I know what it is in software, but never heard of it in hardware terms, outside of ECC memory support there is nothing in Fermi that would not be used for gaming. The same silicon that is used in DP, has two times the single precision, the extra caches actually help performance in many instances, things like adaptive tessellation, where data can be rapidly modified instead of sending it to ram and back.

It was mentioned in the Fermi tech papers when it's GP-GPU capabilities were revealed in late september,but i didn't really check in depth what they meant by that to be honest.


As for dual precision math,it's more of a question of how much transistor budget was spent to give it that much dual precision performance overall,and the same goes for the size of the caches,which when unified,do add extra flexibility such as the case you mentioned,but how much is actually needed in both a gaming scenario and a GP-GPU scenario is my question.


Some users might even ask themselves the question that since Fermi is a processor with a 3 billion transistor budget,how is Cypress keeping up when it's only using 2.15 billion by comparison...That's a 850 million transistor difference afterall,which is a higher amount than the entire transistor budget used for the nvidia G80 GPU for instance(8800 GTX)


This is assuming it's a close fight between Cypress and Fermi on the gaming front of course...People and reviewers are bound to ask themselves that question.

shadow001
03-02-10, 10:43 PM
Interesting... the 470 and 480 both have 384bit buses and 1.5Gb memory. That's unusual for Nvidia.


8800 GTX used a 384 bit bus too,only that it had half as much memory,since memory modules with today's densities didn't exist back then.

Ninja Prime
03-02-10, 10:58 PM
Interesting... http://mylab.q.yesky.com/thread-121-1-1.html

According to this supposed spec sheet leak, the GTX 470 has 448 cores, clocked at 625mhz core 1250mhz shaders, and is only using 3200mhz GDDR5. Kinda defeats the purpose of using bigger buses if you downclock the slowest GDDR5...

In fact, that gives it pretty much the same bandwidth as the 5850.

CaptNKILL
03-02-10, 11:37 PM
8800 GTX used a 384 bit bus too,only that it had half as much memory,since memory modules with today's densities didn't exist back then.

No I meant that both cards had the same bus width according to the boxes. That's unusual for nvidia (8800GTX had 384, 8800GTS had 320, GTX 280 had 512, GTX 260 had 448).

It was a typo on the box though. The 470 will most likely have 320 and the 480 will have 384.

shadow001
03-03-10, 01:00 AM
No I meant that both cards had the same bus width according to the boxes. That's unusual for nvidia (8800GTX had 384, 8800GTS had 320, GTX 280 had 512, GTX 260 had 448).

It was a typo on the box though. The 470 will most likely have 320 and the 480 will have 384.


But up until now with Fermi,the older G80 and GT200 GPU's still used GDDR3,which transfers 2 bits per clock,while GDDR 5 is can transfer 4 bits per clock,so with the same bus,the chip has 2X more bandwith in practical terms when using GDDR 5.

Iruwen
03-03-10, 08:06 AM
Nvidia’s ultra-secure location provided us with probably the most interesting meeting of the day. In one room they had a GTX 480 SLI system up and running, connected to 3 1080p screens powering Need For Speed Shift (in 3D). Next door, a similar system was running Bad Company 2 on the same display configuration. In a third room were numerous boxes running single GTX 480 cards. We are glad to report that the room was not burning down and that it didn’t sound like a wind tunnel which is great... as for the various tech demos on show, each was impressive in its own way. Hair effects on the GTX 480 are significantly ahead of anything we have seen before. The level of detail in PhysX explosions such as the collapsing bridge was also highly impressive.

As for the cards themselves, we are not allowed to show any images today but we will say that reports of massive or loud cooling solutions are well off the mark.

http://www.hardwareheaven.com/articles.php?articleid=142#ixzz0h71F11L9

Behind closed doors Zotac were keen to talk about the upcoming GTX 480 as well as show off their version. Due to NDAs we can’t discuss this

http://www.hardwareheaven.com/articles.php?articleid=142#ixzz0h72hIhTT

Iruwen
03-03-10, 08:12 AM
http://img19.imageshack.us/img19/2191/92410822.jpg


A quick pixel counting exercise makes it ~500mm/sq, using the width of the longest (second) section of PCIe x16 connector as a reference. I make that 71.65mm according to wikipedia, so ~0.746mm/pixel (96 pixels wide), and I make the die area to be 30x30 pixels.

CaptNKILL
03-03-10, 08:42 AM
But up until now with Fermi,the older G80 and GT200 GPU's still used GDDR3,which transfers 2 bits per clock,while GDDR 5 is can transfer 4 bits per clock,so with the same bus,the chip has 2X more bandwith in practical terms when using GDDR 5.

I can't tell if you're trying to argue with me about something or if you're just posting that for the sake of posting it. :p

Razor1
03-03-10, 09:04 AM
It was mentioned in the Fermi tech papers when it's GP-GPU capabilities were revealed in late september,but i didn't really check in depth what they meant by that to be honest.


As for dual precision math,it's more of a question of how much transistor budget was spent to give it that much dual precision performance overall,and the same goes for the size of the caches,which when unified,do add extra flexibility such as the case you mentioned,but how much is actually needed in both a gaming scenario and a GP-GPU scenario is my question.


Some users might even ask themselves the question that since Fermi is a processor with a 3 billion transistor budget,how is Cypress keeping up when it's only using 2.15 billion by comparison...That's a 850 million transistor difference afterall,which is a higher amount than the entire transistor budget used for the nvidia G80 GPU for instance(8800 GTX)


This is assuming it's a close fight between Cypress and Fermi on the gaming front of course...People and reviewers are bound to ask themselves that question.


That is true we have to wait to see what the results are, if it comes out with a 30% advantage or more in Dx 11 games over Cypress, are those transistors wasted? If it can keep up with the HD5970 half the time are those transistors wasted?

Vardant
03-03-10, 10:09 AM
GTX 470 ?


http://img.photobucket.com/albums/v455/lixianglover/d276985e.jpg

Revs
03-03-10, 10:12 AM
You got a link?

Vardant
03-03-10, 10:24 AM
http://we.pcinlife.com/thread-1366899-1-1.html

shadow001
03-03-10, 10:44 AM
That is true we have to wait to see what the results are, if it comes out with a 30% advantage or more in Dx 11 games over Cypress, are those transistors wasted? If it can keep up with the HD5970 half the time are those transistors wasted?

That's the million dollar question basically,and it'll give a solid indication which architecture is more efficient in terms of overall transistor budget/performance and power consumption/performance,at least when it comes to gaming scenarios.


The GTX470 results shown above,if accurate,aren't confidence inspiring that Fermi will have a consistent 30% lead over Cypress,much less come close to the HD5970 cards,as it only has 14% more shader and 20% more memory bandwith than the GTX470.

Iruwen
03-03-10, 10:50 AM
GTX 470 ?


http://img.photobucket.com/albums/v455/lixianglover/d276985e.jpg

Almost no performance loss when switching from 4xMSAA to 8xMSAA, there must be something wrong.

Fotis
03-03-10, 10:51 AM
That looks pretty bad if true! Then again results might be fake.

Vardant
03-03-10, 11:08 AM
Let's look at these numbers from another perspective.

Crysis 4xAA HD 5870 18% faster, 8xAA 0.1% faster :D

Performance hit from 4xAA to 8xAA: 29% on ATI, 11% on NV

DiRT 2 8xAA HD 5870 10% faster

Unigine GTX 470? 8% faster

If this is a 448SP card, as some sources started suggesting lately, I don't think that's bad at all. With some driver magic in the next few weeks, it could get pretty close to HD 5870 peformance across the board.

Fotis
03-03-10, 11:19 AM
According to these number the GTX470 is slower ranging from 10-20% in games.

Judging from the specs GTX480 will be 15-20% faster than GTX470 so the situation looks dangerously close to what Charlie was saying about GTX480 trading blows with 5870 of barely faster.

Then again these numbers might be fake and I'm talking nonsense.:)

shadow001
03-03-10, 11:54 AM
According to these number the GTX470 is slower ranging from 10-20% in games.

Judging from the specs GTX480 will be 15-20% faster than GTX470 so the situation looks dangerously close to what Charlie was saying about GTX480 trading blows with 5870 of barely faster.

Then again these numbers might be fake and I'm talking nonsense.:)


Guru 3D said the same thing too...That it would be a competitive fight between both the GTX480 and HD5870,and the GTX470 sitting somewhere between the HD5850 and the HD5870 performance wise,so if these numbers are accurate,it's no wonder that Nvidia is keeping performance numbers a secret till the last moment possible.


Some people waited 6 months for this card,and it looks like what they waited for is falling short of the ATI asskicker they were hoping for.

hell_of_doom227
03-03-10, 12:04 PM
Guru 3D said the same thing too...That it would be a competitive fight between both the GTX480 and HD5870,and the GTX470 sitting somewhere between the HD5850 and the HD5870 performance wise,so if these numbers are accurate,it's no wonder that Nvidia is keeping performance numbers a secret till the last moment possible.


Some people waited 6 months for this card,and it looks like what they waited for is falling short of the ATI asskicker they were hoping for.

Even if it's 5% faster then HD5870 CCC/Drivers crappola is not worth owning AMD card.

scubes
03-03-10, 12:08 PM
what a stroke nvidia is pulling here i for one im as dissapointed as a rubber duck with no beek.... no wonder they were trying to keep fermi a secret for so long on the benchmarks front.. SHAME ON YOU NVIDIA.....

hell_of_doom227
03-03-10, 12:13 PM
True numbers are is that 470 model sits between 5870 and 5890 and 480 is fast as ATI 5890 (of course when crossfire works properly) which makes AMD offering to look like joke.

scubes
03-03-10, 12:19 PM
i cant wait to see the benchies for the 480 model which is the one i want....

scubes
03-03-10, 12:20 PM
how long do you guys think before we will see some 480 benchies?

shadow001
03-03-10, 12:27 PM
Even if it's 5% faster then HD5870 CCC/Drivers crappola is not worth owning AMD card.


Even if it's 6 months late?
Even if it uses a lot more power?
Even if it's more expensive?
Even if it doesn't really beat the competition?
Even if Nvidia at one point made a statement that Fermi would clearly beat the HD5870 cards?


Shows that some will buy anything that Nvidia makes no matter what it is,and that's sad to see their ability to see things objectively and for they are is long gone.


Remember that reference i made to larrabee in a previous post,and that performance wise,it was only competing with the previous generation products from ATI and Nvidia,and that Intel admitted they were very late with the product,and decided to not release it in it's current form?


That's a 50 billion dollar company saying they missed their launch window,which i stated is crucial on many posts now,and would rather not release a product that doesn't live to performance expectations relative to the current competition,even though this company is the largest one in the industry,and could spend huge amounts of money on a fancy marketing campaign,well beyond the means of ATI or Nvidia.


My take on it?....Nvidia should just keep Fermi for the GP-GPU market and professional 3D rendering markets and not bother with a Geforce version with this initial version,which it's birth was obviously very painfull and delayed.