The GT300/Fermi Thread - Part 2!

[ajaidev]

new final/performance numbers are in:

gtx470 battles 57xx radeon series
gtx480 battles 58xx radeon series
gtx490 battles 59xx radeon series

and this is the best nvidia can for 2010

you guys were right all along nvidia sux!!

hahah first you broke Ati fan boys ballz now you are after Nvidia fan boys...

Geeez you are on a ballz breaking spree

[NapalmV5]

hahah first you broke Ati fan boys ballz now you are after Nvidia fan boys...

Geeez you are on a ballz breaking spree

my loyalty lies with neither

[saaya]

Several of my contacts there (marketing, development, etc) have risen through the ranks over the years I have been talking to them so I don't know where you gotyour info from.

forgot the name of the website... its the no1 site where employees rate and review the companies they worked or work for, plus they give their ceos approval ratings.

As you increase the performance of a GPU, the CPU naturally needs to be faster to feed it information at a quick pace. However, as these GPUs quickly outpace game development, the CPU will continue to be a bottleneck all the way into high instances of AA. Luckily, it seems like DX11 has moved less emphasis off of the CPU which bodes well for the future.

yeah i know... so what do you think? will fermi need a fast cpu or not? they showed it off with a 960...

This really depends on how good of a heatsink NVIDIA sticks on it. I have a feeling though that the combination of 8-pin / 6 pin power connectors and the PCI-E 2.0 slot will be able to provide more power than even an overclocked GF100 can ask. Naturally, when you get into areas like voltage tweaks the consumption of any component will skyrocket.

well if the tdp numbers i heard are true, then its 50W more than a 285... and thats really a lot... i cant imagine what kind of a heatsink that occupies 2 slots is needed to keep that cool... i just wondered if that was only early silicon and if the newer stuff is running cooler...

No it isn't. With AFR each frame is handled completley by either GPU1 or GPUn.

And since the memory is mirrored for both GPUs, you effectively have the same amount, speed, bus width and bandwidth of the single GPU counterpart.

thats not true, while a frame gets rendered there is constantly data written to and read from the mem... and that is NOT mirrored between the two gpus... otherwise both frames would end up identical...

both gpus get the same raw data, i guess, but they then use their memory and memory bandwidth independantly... if they would really mirror each others memory then you would have to split the memory into 2 partitions and the effective memory per gpu would actually drop in half
but why would you do that? why does gpu1 need to know what that gpu is doing with the data and what its frame will look like?

It's not just the consumers, also the shop managers often have no clue about various brands and models of hardware. Whenever I see one of them 'helping' one of the customers... I really wish I wouldn't hear what they were suggesting.
So no wonder renaming works for Nvidia so well.
Most people only look at the numbers... model number, price, VRAM amount, that's it. And a lot assume that the more they pay the better product they get... So there is a lot of room for ridiculous pricing, and people are more than happy to buy such stuff.

its even worse, ive seen several sales people in shops telling people marketing nonsense i could SEE they knew was not true... but they dont care, they want to sell their stuff... i can understand it, but i wouldnt do it...

There are little to no constructive posts found in this thread, hence my post in my opinion is valid. It's just the same old same old rehashed over and over again. Now it's even more of a load of rubbish with even less info.

Forums would exist, and they would be a lot more interesting if people would post some more 'useful' things.

Every day I check this thread and it's bla bla this bla bla that. It's probably because Fermi is so late that I just can't bear this stuff anymore, usually it isn't as bad and actually quite entertaining sometimes, because there is actually stuff to discuss. Now there is nothing much to discuss anymore, and people still keep yapping.

idk, i consider this an offline chat... as soon as something interesting is discussed i go back a page or two to catch up on what happened... i prefer too much info over not enough info that somebody thought was not important... and besides, even if there is no or little info, its fun to talk to others about tech, the companies that make them, their products...

[ElSel10]

thats not true, while a frame gets rendered there is constantly data written to and read from the mem... and that is NOT mirrored between the two gpus... otherwise both frames would end up identical...

Every frame that is rendered using AFR can only use the amount of memory on one card.

The graphics memory is NOT doubled in SLI mode. If you have two 128MB graphics cards, you do NOT have an effective 256MB. Most of the games operate in AFR (alternate frame rendering). The first card renders one full frame and then the next card renders the next frame and so on. If you picture SLI working in this way it is easy to see that each frame only has 128MB of memory to work with.

Quoted from Mad Mod Mike on SLIzone.

[saaya]

Every frame that is rendered using AFR can only use the amount of memory on one card. Quoted from Mad Mod Mike on SLIzone.

what does that have to do with memory bandwidth?
i never said that you end up with double the memory, but you do end up with double the bandwidth from my understanding...

at the same time a dual gpu card is working on two frames, and each gpu can read and write independently to its own memory while working on those frames. as a result, in the same period of time, you end up with (up to) double the frames being rendered, and (up to) double the reads and writes to memory. just think about it... you cant produce additional frames without additional reads/writes to memory...
and think about the real world performance of dualgpu cards... if you would just double the shaders i dont think we could be able to get as much of a performance boost as we see from single to dualgpu cards

what your saying is that both gpus only use the memory of one of the two cards... which makes no sense... rendering a frame takes several steps, you read from memory, manipulate the values and write back to memory... as far as i know its impossible to render two different frames if you force the memory of both gpus to be 100% identical at all times... if you would do that then youd end up with 2 identical frames...

[trinibwoy]

Every frame that is rendered using AFR can only use the amount of memory on one card.

saaya is correct. Capacity is not doubled (both cards hold the same data). However bandwidth is doubled (both cards work on different frames in parallel).

[Tim]

idk, i consider this an offline chat... as soon as something interesting is discussed i go back a page or two to catch up on what happened... i prefer too much info over not enough info that somebody thought was not important... and besides, even if there is no or little info, its fun to talk to others about tech, the companies that make them, their products...

I suppose one can see it like that as well.

[LordEC911]

someone whom I trust told me that the clocks would be GTX 280-like, if a bit higher. So no 700 for you, prolly around 650 like we first thought

Correct, that is what I heard quite awhile ago. B1 is whole new territory though.

I'm not sure what to say about triangles. Over at B3d, they brought up the fact that 2 x 5770's beats 1 x 5870 in triangle intensive games like HAWX, which the GF100 was benched at as being pretty fast. Supposedly 2 cards means twice the tri/clock?

Correct. Each card has it's own setup engine and can effectively double the tri/clock.

[saaya]

saaya is correct. Capacity is not doubled (both cards hold the same data). However bandwidth is doubled (both cards work on different frames in parallel).

so for dualgpu solutions the shader power is doubled, the triangle setup is doubled, texturing power is doubled, memory bandwidth is doubled... but you need double the memory compared to a single gpu to have the same effective memory capacity. and another downside is that you need more cpu power and loss of efficiency when aligning both gpus to work on the same scene...

what i wonder about is that the cpu requirements for dual gpu setups are not double of what a single gpu setup requires. how come?
its definately higher, but not double, at least not in most scenarios... does anybody know why?

no news on fermi tdps?

[annihilat0r]

what i wonder about is that the cpu requirements for dual gpu setups are not double of what a single gpu setup requires. how come?
its definately higher, but not double, at least not in most scenarios... does anybody know why?

Why should it be double? If with say, a 5870 the CPU was a bottleneck, then it would require double power when you add a second 5870. In this age of console games 5870 can be bottlenecked by the CPU quite often, I accept that but when you're bound by the CPU you are at 100 FPS levels so you wouldn't plug in a second card anyway.

As long as the limiting factor is the graphics card I don't think a second card would require double CPU power.

[highoctane]

GTX480 to debut at Cebit: http://www.fudzilla.com/content/view/17586/1/

[saaya]

Why should it be double? If with say, a 5870 the CPU was a bottleneck, then it would require double power when you add a second 5870. In this age of console games 5870 can be bottlenecked by the CPU quite often, I accept that but when you're bound by the CPU you are at 100 FPS levels so you wouldn't plug in a second card anyway.

As long as the limiting factor is the graphics card I don't think a second card would require double CPU power.

mhhh good point, cpu limitation shows slowly, first your only losing 1% of performance then more and more the slower the cpu is compared to the gpu...

GTX480 to debut at Cebit: http://www.fudzilla.com/content/view/17586/1/

cool, ill be there

[zalbard]

cool, ill be there

Will you make cool videos and photos again?!

[saaya]

Will you make cool videos and photos again?!

if i feel like it... maybe
im going there on my own, no sponsor... too bad actually :P

[pingpong1455]

i broke i ended up with a 5870 couldant wait any longer

[Jowy Atreides]

More G92!
The 9600gso 55nm cards with a new name.

8800gt = 9800gt = GT 240 = GT 340

identical, equal performance, different names, 2 of which imply there is performance upgrade

[trinibwoy]

Nice. Rise of the Undead Part IV.

[eric66]

die g92 dieeeeeeeeeeeeeeeeee

[LesGrossman]

gtx 580 g92 edition , thank god for the tags.

[570091D]

lol, i don't remember any g92 cards with 768mb of memory.... and gpu-z says it's a gt330 not a gt340.

[GoldenTiger]

lol, i don't remember any g92 cards with 768mb of memory.... and gpu-z says it's a gt330 not a gt340.

The original 9600gso's had 96 stream processors and 768mb or 384mb configurations for RAM providing a 192-bit bus, G92-based. They later gimped them to 48sp with 512mb RAM and a 128-bit bus.

[***Deimos***]

The original 9600gso's had 96 stream processors and 768mb or 384mb configurations for RAM providing a 192-bit bus, G92-based. They later gimped them to 48sp with 512mb RAM and a 128-bit bus.

a lot of gimping going on at nVidia. Its like a Frankenstein GPU, PCB garage sale.

die g92 dieeeeeeeeeeeeeeeeee

look at my quote in Jowy Atreide;s sig.

A history lesson in complacency.
Like no other industry in the history of the world, computers ushered in dramatic increases in performance and functionality and unheard of price reductions. Competition is fierce. Those who take a break, and fail to push the boundaries are doomed to be amongst the forgotten has-beens: Cyrix, 3Dfx, VIA, S3, Abit.

4 years of milking Athlon64/Opteron sales, and a delayed Barcelona with TLB Bug almost crushed AMD.

That's why nVidia's 2007-2010 rebrandfest is concerning. Sure, way back when before 8800GT, you could argue that DX10.1 was a novelty. But time goes by fast. A hush-hush DX10.1 GT240 rollout, 2 MONTHS AFTER AMD launched DX11 cards... pathetic. Just because you were making money yesterday, doesn't guarantee future revenue.

Its a mystery that nVidia alone has taken upon themselves to sabotage graphics progress. Its time to get act together. Optimus is a great start and should be in EVERY notebook.

No more 5-7 month delays for launch of DX11 Fermi mainstream and value derivatives. Bad Company 2 is coming out in 20 days. Hup-too-hup-too double time soldier!

[Jowy Atreides]

lol, i don't remember any g92 cards with 768mb of memory.... and gpu-z says it's a gt330 not a gt340.

9600GSO

And the 240 is an 8800gt, 230 is gso

[Chumbucket843]

yeah, they just magically rebranded directX 10.1 into their chips too. 40nm doesnt count for anything either? i guess the definition of rebrand has changed. if thats the case then many chips are just rebrands and no one should buy those. in my opinion if it still uses silicon its a rebrand.

[570091D]

8800gt = 9800gt = GT 240 = GT 340

identical, equal performance, different names, 2 of which imply there is performance upgrade

9600GSO

And the 240 is an 8800gt, 230 is gso

you're contradicting yourself... and how exactly is this GF100 news?

yeah, they just magically rebranded directX 10.1 into their chips too. 40nm doesnt count for anything either? i guess the definition of rebrand has changed. if thats the case then many chips are just rebrands and no one should buy those. in my opinion if it still uses silicon its a rebrand.

Jowy A, shouldn't you be mad at ati for rebranding the 4870 as a 5770? just because it's built on 40nm and has dx11 doesn't mean it's new!