The GT300/Fermi Thread

[Sam_oslo]

Benchmarks change as do which games benefit certain manufacturers. However, FC2 at Ultra High details remains one of the most demanding games in any category.

On the flip side of that coin, this was a PR event by NVIDIA which means they showed what they wanted to show and no more. I don't think average performance will be close to what was shown but then again, considering drivers can mature and clock speeds will increase, you never know...

You are right, and in case we are looking at a single GPU with is going to match the HD5970. Then a GTX 380 with 50-60% over HD5870 sounds just about right!
I was more focused on HD5870, and was taking about the GTX 360, but I should persist that. Sorry for a bad formulated post and cousin misunderstandings.

[Solus Corvus]

Sure I can. IMO, there is something wrong with their 2560 result. Why? Take this into consideration:

1680 -> 1920 = 540,000 pixels difference = 7.1 FPS drop

1920 -> 2560 = 1,792,000 pixels difference = 3.4 FPS drop


Their 1920 x 1200 results look good.

Rejecting their 2560x1600 numbers because they don't meet your expectations and accepting their 1920x1200 numbers because they do is not sound reasoning. If their methodology or test equipment is suspect then all their numbers are suspect.

http://www.techpowerup.com/reviews/A...D_5870/14.html

Then a GTX 380 with 50-60% over HD5870 sounds just about right!

In the only game benchmark shown it's 31-39% faster then a 5870, not 50-60%.

[Sam_oslo]

...
In the only game benchmark shown it's 31-39% faster then a 5870, not 50-60%.

But my calculations shows GTX 380 needs to perform 50-60% over HD5870 to match the HD5970. The GTX 360 would do with 20% over HD5870. This would be the best business scenario for nVidia, I believe, and I bet they would give it a go.

I agree, it sounds like a impossible task, specially after all those speculation and bashing some people () have been trowing around this forum, lately. But maybe they can get there with matured drivers?

[Solus Corvus]

If the numbers shown are from a 448 shader card then the full 512 shader cards could match a 5970. But if the numbers are from the full shader card then they won't be able to match a 5970 unless they make a dual-gpu card.

[Sam_oslo]

Yeah that can be one scenario, too.
But even if these numbers are from GTX 380 (512 shader), considering that these are using very early drivers, and these are totally new driver, don't you think they got a good chance to smack over the HD5870 by 50-60% by lunch time?

[Solus Corvus]

No, I don't.

[zerazax]

Yeah that can be one scenario, too.
But even if these numbers are from GTX 380 (512 shader), considering that these are using very early drivers, and these are totally new driver, don't you think they got a good chance to smack over the HD5870 by 50-60% by lunch time?

2 months worth of drivers will not make a card go up 30% in overall performance. You'd have to have the world's worst drivers right now. That's just wishful thinking at best

And from most indications, that was the GTX 380 full core being demo'd

[ElSel10]

But my calculations shows GTX 380 needs to perform 50-60% over HD5870 to match the HD5970.

Why is that? from what I've seen the 5970 is about 40% faster than the 5870, and that is average FPS, not minimum.

[highoctane]

Just going from a core count standpoint a 360 should be capable trading blows with a 5870 and the 380 is probably what was being demo'd, I would think. If the demo was say an A2 silicon lower clocked 380 512 core based unit then the A3 respin may allow for better clock binning and possibly better performance by launch time.

At any rate the cards should be fast but I'm still waiting for some solid reviews...

[eric66]

why you guys don't wonder why nvidia didn't bench dirt or stalker or other dx 11 games to show there great tessellation power. i mean if it can run fc bench then how about trying fermi in rel dx 11 games instead of homemade demos

[annihilat0r]

why you guys don't wonder why nvidia didn't bench dirt or stalker or other dx 11 games to show there great tessellation power. i mean if it can run fc bench then how about trying fermi in rel dx 11 games instead of homemade demos

My understanding is that Nvidia has aimed big, which has a lot of advantages and disadvantages. Now, when Nvidia does a presentation about this before the cards are out, they are obviously going to point out the advantages (tessellation, geometric computing, GPGPU etc.) and leave out the disadvantages (price, heat). And from what we are hearing, 1.25x 5870 is what Fermi is going to be, and that isn't exactly a great advantage.

So Nvidia leaves that out. Dirt and Stalker don't use enough tessellation to make Fermi really stand out, so it's going to be at best 1.3x better than a 5870.

[zerazax]

My understanding is that Nvidia has aimed big, which has a lot of advantages and disadvantages. Now, when Nvidia does a presentation about this before the cards are out, they are obviously going to point out the advantages (tessellation, geometric computing, GPGPU etc.) and leave out the disadvantages (price, heat). And from what we are hearing, 1.25x 5870 is what Fermi is going to be, and that isn't exactly a great advantage.

So Nvidia leaves that out. Dirt and Stalker don't use enough tessellation to make Fermi really stand out, so it's going to be at best 1.3x better than a 5870.

This is a good point for everyone to remember. This was a PR presentation. They'll highlight the good and keep the bad under covers. After the hype wears off, read between the lines.

For example... They said 2.33x faster for 8xAA than a GTX285. That's great... Until one remembers that the GT200 and G92 cards were notorious for choking at 8xAA. I have no doubt Fermi will be the fastest single GPU but I'm mindful of the PR work at it's best

[LordEC911]

Just going from a core count standpoint a 360 should be capable trading blows with a 5870 and the 380 is probably what was being demo'd, I would think. If the demo was say an A2 silicon lower clocked 380 512 core based unit then the A3 respin may allow for better clock binning and possibly better performance by launch time.

At any rate the cards should be fast but I'm still waiting for some solid reviews...

Why would they use A2? Since they left out clocks we don't know for sure but this was either A3 clocks, aka likely launch clocks, or their target clocks, which they seem unlikely to hit.

There is talk about a B1 in the works. Launch A3, give what few full speed and full shader parts to reviewers, then once B1 is ready just use those parts. Don't know how likely that is but it was hinted at from a Chinese forumite.

[orangekiwii]

Someone said it best a couple pages back, but I admire fermi much more (even if it performs similarly to 5870) than the 5xxx series.


What games REALLY need the horsepower this brings? Only multi-monitor displays with massive resolutions with heaps of AA/AF you won't notice.

While its not 'shown' (demonstrated with independent reviews) it appears that fermi will dominate in "next gen" areas (tessalation, geometry, particle simulation...).

so basically 5870 = fermi in current/old games but fermi > 5870 in next gen games (of course next gen means non console remakes so this is all moot by the time we get non-consolized games [never?] much better hardware will be out)


I think fermi will serve nvidia incredibly well as if it is indeed "future proofing itself" then tweaks to that and improvements on classic rasterization will just make the architecture best of both worlds. I highly suspect 6xxx from AMD will turn out to be extremely similar to fermi in design principle (maybe not size but the way it handles information)

[annihilat0r]

This is a good point for everyone to remember. This was a PR presentation. They'll highlight the good and keep the bad under covers. After the hype wears off, read between the lines.

For example... They said 2.33x faster for 8xAA than a GTX285. That's great... Until one remembers that the GT200 and G92 cards were notorious for choking at 8xAA. I have no doubt Fermi will be the fastest single GPU but I'm mindful of the PR work at it's best

I think Nvidia is giving hints of real Fermi performance in that 4x-8xAA scenarios example in Hawx.

Compare 4xAA 285 to 4xAA Fermi. It's a 60 percent increase.

[LordEC911]

I highly suspect 6xxx from AMD will turn out to be extremely similar to fermi in design principle (maybe not size but the way it handles information)

It has been hinted at for awhile, since before Cypress, that a new AMD/ATi architecture would be able to handle more than 1tri/clock, though that might have been due to Cypress' "dual rasterizer."
We can only hope that with their new architecture that they do up the ante in that area.

[Andrew LB]

I can give you for a fact that MP doesn't start for at least another month.

A fact? Really?

Two weeks ago I received my monthly investor update from TSMC which noted mass production has begun on nVidia's high end 40-nm GPU after months of delays caused by wafer production problems.

nVidia's CEO also clearly stated at CES 2010 that the GF100 is now in mass production.


Now I'm sure some of you out there are going to say "oh well.... yea.... nVidia lies all the time".

See.... if it were just the comment @ CES, I'd probably agree with you guys that nVidia is just trying to string buyers along so they don't buy an ATi card. But some people forget that it is HIGHLY ILLEGAL to lie in shareholder reports. In addition, why would TSMC break the law at the request of nVidia?

TSMC did over $10 Billion in sales in 2008 and nVidia is only a tiny part of that

[JohnZS]

I have heard that it is all gearing up for wide availability of Fermi in March.
Instead of doing an Ati and releasing the product early with insufficient drivers and poor availability (HD 58xx series cards did not come out in the UK until December/Last Week of November at the earliest and HD5970 cards are still not released here) nVidia will be releasing on mass in March with relatively mature drivers too (well I do say relatively, obviously there would still be the odd gremlin or two that you get with a new card for the first couple of months).
John

[annihilat0r]

There's a difference between mass production of GPUs and mass production of cards. The second can't start before the first is finished. So I guess neliz is talking about the production of cards, i.e. the placement of GPUs on the boards; while TSMC is obviously talking about production of the dies, because well, it doesn't have anything else to do.

[Florinmocanu]

I have heard that it is all gearing up for wide availability of Fermi in March.
Instead of doing an Ati and releasing the product early with insufficient drivers and poor availability (HD 58xx series cards did not come out in the UK until December/Last Week of November at the earliest and HD5970 cards are still not released here) nVidia will be releasing on mass in March with relatively mature drivers too (well I do say relatively, obviously there would still be the odd gremlin or two that you get with a new card for the first couple of months).
John

Dude, get over England and the fact the cards where launched later on.

In Europe, US, Asia all 5xxx GPUS launched at the correct date, so Uk's problem is a local one, not a general one.

Problems with stocks exist and have existed, but you still could buy one, lots of people bought 5xxx series.

And stop a_ss kissing nvidia "because they don't release the product cause it has problems with availability", because they don't have it yet ready, that's why they are not releasing it. When they will have final cards in their hands, they will do a paper lauch, just like with GTX 295, don't worry about that.

[Jowy Atreides]

Instead of doing an Ati and releasing the product early with insufficient drivers and poor availability (HD 58xx series cards did not come out in the UK until December/Last Week of November at the earliest and HD5970 cards are still not released here)

...Actually....

I got mine launch day. In the UK.

Go figure.

PS. the drivers were fine and still are

[Blacky]

[FUD] Dual GF100 Fermi should be in April ohhhh myyyyy gawwd

[annihilat0r]

Dual Fermi isn't going to be much faster than a single full Fermi, it's only being released because a single GF100 isn't enough to get past 5970. Dual Fermi will probably be faster than a 5970 by the minimum amount it would be required to call it a faster card than 5970. The preceding sentence rocked, right! Display of my elite English skillz.

[Blacky]

Dual Fermi isn't going to be much faster than a single full Fermi, it's only being released because a single GF100 isn't enough to get past 5970. Dual Fermi will probably be faster than a 5970 by the minimum amount it would be required to call it a faster card than 5970. The preceding sentence rocked, right! Display of my elite English skillz.

I hope you can hook up three displays with one of these

[W1zzard]

Dual Fermi isn't going to be much faster than a single full Fermi

uh why?