The Fermi Thread - Part 3

[tajoh111]

Ummm... architecture is ~6months old...
Just because it isn't a complete architectural overhaul doesn't mean a lot of things didn't change under the hood that takes time to optimize. Look at all the new code that is being used thanks to a new API.

To think that AMD/ATi could have been optimizing and coding their drivers better for a future (unreleased) product is completely inane.
.

I think you missed my point. I am saying we can't expect great performance from fermi when it was obviously rushed to the market and little time was available to optimize drivers. You of all people who has been a strong charlie supporter, thinking NV didn't have working silicon till somewhat recent times.

Since this was response to someone who said they had for months to optimize drivers for fermi so we should expect great performance off the bat, if your saying their going to be work being done on AMD hardware because there are still some changes(hence the 6-7 month time frame), well on something completely new as in fermi, your going to have to do alot more work; atleast more work than what was needed to get the performance jumps on the 5xxx series anyways(which still took months to get).

The thing is current drivers for for anything AMD supports everything based on 2900xt and up, which is why its far easier for AMD to get gains since they are all based on the same architecture or r600.

Fermi is completely new but shares drivers that are primarily for g80 based architecture. As a result of this specialization, its harder to make drivers that give fermi based cards a boost while not hurting performances of cards from g80 to g200. Some of the biggest gains from AMD drivers when they decided to abandone driver support for 19xx and older generation and the reasons are obvious.

The amount of work to get fermi drivers up to its maximum potential is likely going to be one of the most complex driver projects out there because they still have to cater to the architecture before fermi and the complexity of fermi itself. We both know this, hence a 4 month time to write drivers that wring out alot of performance from fermi is not enough, it might take years. Ideally the plans to skip should have the real 40nm g200 refresh should have come sooner, as fermi in its current state is a beta project with little driver development. The potential for fermi to perform is really there as some benchmarks show as its biggest gains are in programs which have been in the past AMD friendly.

[Sn0wm@n]

anyone else find this odd that R600 got beaten down by g80 ... yet a couple years down the line and its giving a fair fight to nvidia's next gen arch

[foch3 -USA-]

Ringbus?

[annihilat0r]

anyone else find this odd that R600 got beaten down by g80 ... yet a couple years down the line and its giving a fair fight to nvidia's next gen arch

means it was a scalable architecture

[zerazax]

anyone else find this odd that R600 got beaten down by g80 ... yet a couple years down the line and its giving a fair fight to nvidia's next gen arch

means it was a scalable architecture

It really is ironic. R600 was late, hot, unoptimized, slow, etc. - just about everything possible that could have gone wrong, went wrong. People forget... at its release, people were talking about AMD/ATIs bankruptcy, which was a real possibility.

RV770 comes out, which is a bit slower than the GT200s, but comes out at roughly the same time and is shockingly fast for its price and size. RV870 comes out on time and is actually cooler than the previous generation, and manages to hold its own against Nvidia's next gen.

Strangely, out of all the cards now looking back, perf/watt and perf/mm^2 have gone to ATI's cards based on the inferior (at the time R600) whereas Nvidia's just hasn't been able to scale up OR down (we were stuck with G92s for years, by the time the GT200 derivatives came out, the RV870 and its derivatives were already on the doorstep)

[Sn0wm@n]

It really is ironic. R600 was late, hot, unoptimized, slow, etc. - just about everything possible that could have gone wrong, went wrong. People forget... at its release, people were talking about AMD/ATIs bankruptcy, which was a real possibility.

RV770 comes out, which is a bit slower than the GT200s, but comes out at roughly the same time and is shockingly fast for its price and size. RV870 comes out on time and is actually cooler than the previous generation, and manages to hold its own against Nvidia's next gen.

Strangely, out of all the cards now looking back, perf/watt and perf/mm^2 have gone to ATI's cards based on the inferior (at the time R600) whereas Nvidia's just hasn't been able to scale up OR down (we were stuck with G92s for years, by the time the GT200 derivatives came out, the RV870 and its derivatives were already on the doorstep)

ty for understanding the irony

[Behemot]

The thing is current drivers for for anything AMD supports everything based on 2900xt and up, which is why its far easier for AMD to get gains since they are all based on the same architecture or r600.

Fermi is completely new but shares drivers that are primarily for g80 based architecture. As a result of this specialization, its harder to make drivers that give fermi based cards a boost while not hurting performances of cards from g80 to g200. Some of the biggest gains from AMD drivers when they decided to abandone driver support for 19xx and older generation and the reasons are obvious.

This is nice manipulation with facts, you know? If you say Fermi is completely new, than you must say Evergreen is completely new architeture, too. AMD did the major changes with RV7x0 generation (memory hub and other stuff), NVIDIA with GF10x (cache and other stuff).

But the truth is, the shaders and other parts which have direct impact on 3D performance are the same till R600/G80. So this talk about "hard to make drivers especially for GF100" is total nonsense.

If they want to make special driver, nobody is holding 'em. It is so easy to name the DLL libraries differently and copy the different ones if GF100 is detected, isn't it?

[ElSel10]

If you say Fermi is completely new, than you must say Evergreen is completely new architeture, too. AMD did the major changes with RV7x0 generation (memory hub and other stuff), NVIDIA with GF10x (cache and other stuff).

Not at all. RV870 is far more closely related to the RV770 (and hence R600) than Fermi is to GT200 (G80).

It really is ironic. R600 was late, hot, unoptimized, slow, etc. - just about everything possible that could have gone wrong, went wrong. People forget... at its release, people were talking about AMD/ATIs bankruptcy, which was a real possibility.

The only similarities Fermi shares with R600 is that it is hot, power hungry and ~6 months late. The fatal blow to the R600 was that when compared to the competition (G80), it was woefully underpowered and slow, and in fact, the X1900 cards actually outperformed it in certain circumstances when AA was enabled.

That is not the case for Fermi. Yes it's hot, power hungry and late, but it at least maintains the title of fastest single GPU, a big differentiation from R600. Further, AMD was and still is far closer to bankruptcy than Nvidia.

[LordEC911]

I think you missed my point. I am saying we can't expect great performance from fermi when it was obviously rushed to the market and little time was available to optimize drivers. You of all people who has been a strong charlie supporter, thinking NV didn't have working silicon till somewhat recent times.

Since this was response to someone who said they had for months to optimize drivers for fermi so we should expect great performance off the bat, if your saying their going to be work being done on AMD hardware because there are still some changes(hence the 6-7 month time frame), well on something completely new as in fermi, your going to have to do alot more work; atleast more work than what was needed to get the performance jumps on the 5xxx series anyways(which still took months to get).

The thing is current drivers for for anything AMD supports everything based on 2900xt and up, which is why its far easier for AMD to get gains since they are all based on the same architecture or r600.

Fermi is completely new but shares drivers that are primarily for g80 based architecture. As a result of this specialization, its harder to make drivers that give fermi based cards a boost while not hurting performances of cards from g80 to g200. Some of the biggest gains from AMD drivers when they decided to abandone driver support for 19xx and older generation and the reasons are obvious.

The amount of work to get fermi drivers up to its maximum potential is likely going to be one of the most complex driver projects out there because they still have to cater to the architecture before fermi and the complexity of fermi itself. We both know this, hence a 4 month time to write drivers that wring out alot of performance from fermi is not enough, it might take years. Ideally the plans to skip should have the real 40nm g200 refresh should have come sooner, as fermi in its current state is a beta project with little driver development. The potential for fermi to perform is really there as some benchmarks show as its biggest gains are in programs which have been in the past AMD friendly.

So you are saying Nvidia had working GF100 silicon in May? I'm a Charlie supporter when his sources/rumors seem to support info I heard from a different source. Nothing more or less. If you want to look back, I actually disagreed with Charlie quite a few times over the last year or so.

You also missed what I am trying to say... Evergreen is a new architecture, in the sense of drivers, the same Fermi is. Did AMD/ATi have working silicon before their Sept launch, more than likely, did Nvidia have working silicon before the end of 2009, more than likely, I never said any different. The point is, to say that AMD/ATi should have already optimized the performance from Evergreen before it's launch, or shortly therefore after, is wrong. It will take months to correct and optimize, sort of like how it took +6 months for G80 to get stable drivers in certain games(though not the same).

I think the SLI driver performance shows how much time Nvidia really had to prepare GF100. The performance scaling is quite good and somewhat surprising, though that might be only my opinion.

So Nvidia dropping supporting for pre-G200 is of no consequence to you? We will see...

[Behemot]

Not at all. RV870 is far more closely related to the RV770 (and hence R600) than Fermi is to GT200 (G80).

Oh yes, keep repeating. It is closer to RV770, I did not say it isn't, but definitely is far away from R600. If you say so, you mean the gaming side of GPU is the same and in this case it is just right to say the GF100 is same as G80, too. In both architectury, AMD and NVIDIA, is full ton of GPGPU changes from old R600%G80. But who in cares?! You play video encoding? You play Photoshop or what? I'd say 99% play games.

So if the cards are 20% better on average, it'll stay that. No driver-miracles possible. I only hope people will finally end with this "drivers will change everything" BS

[annihilat0r]

Not at all. RV870 is far more closely related to the RV770 (and hence R600) than Fermi is to GT200 (G80).



The only similarities Fermi shares with R600 is that it is hot, power hungry and ~6 months late. The fatal blow to the R600 was that when compared to the competition (G80), it was woefully underpowered and slow, and in fact, the X1900 cards actually outperformed it in certain circumstances when AA was enabled.

That is not the case for Fermi. Yes it's hot, power hungry and late, but it at least maintains the title of fastest single GPU, a big differentiation from R600. Further, AMD was and still is far closer to bankruptcy than Nvidia.

2900XT was slower than G80 but was priced very good, about 8800GTS 640mb I believe, and was a good alternative to it in a number of ways.

GTX 480 might be the "fastest single GPU", but it's by no means the fastest card. It's just a (good) alternative to the 2nd fastest card, like 2900XT was.

There'll definitely be a dual GPU version but how, without a die shrink, I can't understand. It looks like even a B1 revision (which won't happen, according to some people) and increased yields won't allow a <300w dual Fermi. Maybe they'll scratch the 300w limit. Heck, it looks like they've gone over it with the 480 sp single GPU version!

R600 sucked but the architecture proved to be upwards scalable. It appears that Fermi design is downwards scalable (=GF104 and 108 coming), so by logic it should also be upwards scalable And that should make some great chips after node shrinks.

[Humminn55]

I am saying we can't expect great performance from fermi when it was obviously rushed to the market and little time was available to optimize drivers.

I'm sorry, but rushed to market? That made me laugh coffee out my nose.

[annihilat0r]

Oh yes, keep repeating. It is closer to RV770, I did not say it isn't, but definitely is far away from R600. If you say so, you mean the gaming side of GPU is the same and in this case it is just right to say the GF100 is same as G80, too. In both architectury, AMD and NVIDIA, is full ton of GPGPU changes from old R600%G80. But who in cares?! You play video encoding? You play Photoshop or what? I'd say 99% play games.

So if the cards are 20% better on average, it'll stay that. No driver-miracles possible. I only hope people will finally end with this "drivers will change everything" BS

RV770 was released 2 years ago, and by all means RV870 is pretty much a shrank + upscaled RV770. So do you think this is the same situation as GF100 vs GT200b?

Drivers won't change everything but I'm pretty sure GTX 480 will see more driver-based performance improvements in the future than HD5870.

[Macadamia]

2900XT was slower than G80 but was priced very good, about 8800GTS 640mb I believe, and was a good alternative to it in a number of ways.

GTX 480 might be the "fastest single GPU", but it's by no means the fastest card. It's just a (good) alternative to the 2nd fastest card, like 2900XT was.

There'll definitely be a dual GPU version but how, without a die shrink, I can't understand. It looks like even a B1 revision (which won't happen, according to some people) and increased yields won't allow a <300w dual Fermi. Maybe they'll scratch the 300w limit. Heck, it looks like they've gone over it with the 480 sp single GPU version!

R600 sucked but the architecture proved to be upwards scalable. It appears that Fermi design is downwards scalable (=GF104 and 108 coming), so by logic it should also be upwards scalable And that should make some great chips after node shrinks.

IMO Fermi's scalability potential is upwards, not downwards.

We'll see soon, but considering the current ROPs are pretty inefficient (as you can see here Fermi has less fillrate than GTX285), don't expect the GTS 450 to handle 2560, and even in 1920 you might see bad omens.

Junpier aka HD5770 can possibly outclass/come near something more than double its size. But again this time GF104 has more bandwidth- bus width is 2x, not 1.5x. 5850 of the same size? Pretty impossible to beat.


Unless it's 1024*768. Yep, the perfect card for those poor nVidia kids using CRTs from 2000.

[LordEC911]

Junpier aka HD5770 can possibly outclass/come near something more than double its size. But again this time GF104 has more bandwidth- bus width is 2x, not 1.5x. 5850 of the same size? Pretty impossible to beat.

Unless it's 1024*768. Yep, the perfect card for those poor nVidia kids using CRTs from 2000.

GF104 will be almost double Juniper's size.

[Macadamia]

GF104 will be almost double Juniper's size.

Isn't it larger than Cypress? (which is 334)

But blah, semantics. The difference in cost for the same performance will be astounding though. Nowadays you can get a 5770 for 140USD, so I guess even with marginal performance benefits nVidia has to sell their 300squaremm card at $200 max.

I think they've given up on GF106 and proceeded directly to GF108, 64-bit with "effectively" 4 GT200esque ROPs. Brutal.

[Sly Fox]

Man I love you guys. I can go out for the evening, come home, and there's three more pages of ranting for me to read.

Sorry for the OT, been doing a little of the ole'

[Behemot]

RV770 was released 2 years ago, and by all means RV870 is pretty much a shrank + upscaled RV770. So do you think this is the same situation as GF100 vs GT200b?

Drivers won't change everything but I'm pretty sure GTX 480 will see more driver-based performance improvements in the future than HD5870.

No it is not and that's what I tried to point out. If you say the GF100 is different than GT200, you just cannot say RV870 is the same as R600 because they architecturaly differ to. Than if you say R600=RV870, than by this logic you have to say G80=GF100.

By the way, as for terms of shader units, this is true Both firms improvements were for GPGPU mainly. Gaming related is only add-on of new instructions for DX 10.1 and DX 11, adding shader units, memory and some other stuff (like improving compression algorythms, better anti-aliasing etc.).

We'll see, but I personally doubt there won't be some drastical performance speed-ups. The Catalyst 10.2/10.3 impact on speed is speculative, too. These were much more important like bug-fixes and that's what I expect from ForceWare too.

[annihilat0r]

No it is not and that's what I tried to point out. If you say the GF100 is different than GT200, you just cannot say RV870 is the same as R600 because they architecturaly differ to. Than if you say R600=RV870, than by this logic you have to say G80=GF100.

By the way, as for terms of shader units, this is true Both firms improvements were for GPGPU mainly. Gaming related is only add-on of new instructions for DX 10.1 and DX 11, adding shader units, memory and some other stuff (like improving compression algorythms, better anti-aliasing etc.).

We'll see, but I personally doubt there won't be some drastical performance speed-ups. The Catalyst 10.2/10.3 impact on speed is speculative, too. These were much more important like bug-fixes and that's what I expect from ForceWare too.

But then, if gaming-related enhancements and differences of GF100 are adding shader units and stuff, how do you account for the incredible performance differences between games? One game it tramples 5870, one game it falls behind with 2x transistors.

Some guys at B3D have pointed out that this is related to how geometry intensive a game is. Apparenetly FC2 is such a game whereas Crysis is just shader/texturing intensive, thus accounting for the big performance difference.

And if this is true, it should mean that GF100 is architecturally very different.

[Behemot]

But then, if gaming-related enhancements and differences of GF100 are adding shader units and stuff, how do you account for the incredible performance differences between games? One game it tramples 5870, one game it falls behind with 2x transistors.

Some guys at B3D have pointed out that this is related to how geometry intensive a game is. Apparenetly FC2 is such a game whereas Crysis is just shader/texturing intensive, thus accounting for the big performance difference.

And if this is true, it should mean that GF100 is architecturally very different.

This is related to how the game is written. Most performance-increases could be achieved by optimizing the game, no optimizing the drivers. This is too much overestimated way, but wrong way and for that reason we see mostly few percent increases.

I give Unreal Engine as an example every time: as Epic presented on GDF 2008, they focus on optimizing the engine/games before it is released, because it is cheaper. They e.g. have several machines running the game 24/7 and logging it. When somewhere is performance drop, they focus on the reason. We all know Bioshock looks pretty good and runs smoothly on almost everything DX10 capable (if you not count GF8400/HD 3450 as a DX10 capable). This is probably last developer doing this, others usually roll it out with critical bugs and don't care. But it pays off since UE is the most common engine.

And of course this means also how the engine works. For example old games often used post-proces rather than advanced shader features, so they do not perform much better on Radeon X800 or HD 3870. F.E.A.R. is example I remember since I play it when I have time now (once a week or so); I am used to play on 20 FPS average, but this game sometimes falls under 20 FPS even with HD 3870@915 MHz, Windsor 2,8 GHz.

You should also not forget on how AMD's stream processors work: if the SP are feeded optimaly, up to 5 instructions could be processed. If not, you gain performance drop. Than it is huge difference if you feed 3 or 5 instruction on average making the Radeons somewhere win, somewhere lose.

[mAJORD]

I seem to have missed it along the line since NDA lifted but what is the official GF100 die size?

[ajaidev]

For the MIN FPS argument please read Bit techs review

As you can see other than Dirt 2 and Bad Company 2 "When AA is active" the MIN FPS difference is really small among GTX 480 and 5870....!!

[LinkinParkBoy]

nVidia flop.

HD Radeon 5870 + GT240 (PhysX) = Priceless

[Andrea deluxe]

from hardware canucks:

In our tests, thermal throttling usually set in at around 105°C on the GF100 cards while complete thermal shutdown happens around 112-125°C


is real???

[onethreehill]

i seem to have missed it along the line since nda lifted but what is the official gf100 die size?

530 iirc