NVIDIA GF100 & Friends speculation

[Ailuros]

I think it's typical of all GPUs. Cypress' TDP is 188W, yet average load power consumption is around 120-140W (source. And yes, under FurMark, virtually every card exceeds their TDPs.

The average load power of the GTX295 though is at 181W (with a 6pin+8pin, 289W TDP), which might be in line with GF100's TDP. The latter's average load power better not be on roughly 295 level, otherwise I sense quite some headbanging with something short of 50% higher power consumption.

 

[Silus]

Power consumption close to 5970, performance closer to 5870 outside some artifical benchmarks.

If the "30% faster than Cypress" holds up, performance will be closer to 5970.
Power consumption at load may be closer to 5970, but at idle I doubt it will be too far off the 5870.

 

[neliz]

Where'd he go BTW? He was supposed to give me an attributed quote from NVIDIA saying they would ship in '09.

Got his post: http://forum.beyond3d.com/showpost.php?p=1343414&postcount=427

Fudo and i were both there with Drew Henry and Jason Paul (my company is AlienBabelTech [no links posted to there by me]; not Fudzilla)
- they said what Jensen showed "is a production mock-up"

However, they have working silicon of Fermi ... is in their testing setups 'black box', so it is not really "presentable" in a finished videocard

Who really cares if it was a "mock up" or not ?
- Fermi is real and according to the aforementioned Nvidia officials - "It will ship this year" (2009)

 

[MfA]

Interesting, so they really thought it was going to ship ... would be nice if Fudo or Apoppin asked them what happened after the Fermi launch has blown over and they can talk a little more freely (the journalistic thing to do).

 

[trinibwoy]

He decided to give me some extra help over at XS with all the BS/non-info.

Not sure if you mean he's helping you spread the BS (sorry if I misunderstood). I'm waiting with baited breath to hear more about these disabled TMUs that he keeps telling everyone about. Should be a blast.

"Artificial benchmarks" meaning what? This sounds like an attempt to downplay a competitor's tessellation performance.

It's ironic really. You would think AMD would be the one busting out the gate with high tessellation/geometry performance given they're on their 3rd or 4th implementation. But I guess that could be coming soon in their next overhaul.

 

[Ailuros]

It's ironic really. You would think AMD would be the one busting out the gate with high tessellation/geometry performance given they're on their 3rd or 4th implementation. But I guess that could be coming soon in their next overhaul.

Just for the record I protested strongly that a popular tesselation benchmark should be rated as something like an industry standard in the past. Guess what I still do for the same reasons. I'm just not that sure that the gentlemen who made the original suggestion has still the same opinion....*shrug*

 

[dnavas]

Wasn't there a rumor going around that Fermi was 5870+130% (5870x2.3) in Unigine?

The great thing about the rumors right now is that there are rumors that will support nearly any conclusion.

Perhaps this is one of those scenarios.
If seahawk is right, it would be quite bitter. Better quality fps due to the lack of AFR but perf/watt worse than 5870...ouch.

I would have a hard time believing that tesselation is such a key gating factor that a reasonably faster implementation makes a difference of 130% to a final demo. Amazing if so.

<80 filtering units isn't necessarily a problem if you've taken a TMU, split the fetching from the filtering, and clocked the filtering hardware at the rate of the rest of your ALUs. Absent that, I would worry about trilinear "shortcuts"

 

[Mize]

I have it on good authority that Fermi rocks Doom and Q3A...

 

[Sontin]

GF100 in action?!
http://vimeo.com/8754848
http://vimeo.com/8754676
http://vimeo.com/8754728
http://vimeo.com/8754813
http://vimeo.com/8754874

 

[trinibwoy]

Just for the record I protested strongly that a popular tesselation benchmark should be rated as something like an industry standard in the past. Guess what I still do for the same reasons. I'm just not that sure that the gentlemen who made the original suggestion has still the same opinion....*shrug*

Well I can only think of one (Unigine) but that's far from even 3dmark level given that the tessellation usage there wasn't very practical, or pretty IMO.

I would have a hard time believing that tesselation is such a key gating factor that a reasonably faster implementation makes a difference of 130% to a final demo. Amazing if so.

In an actual game? No way. But I fully expect one of 3dmark's "feature tests" to feature simply shaded but highly tessellated model(s). In that situation you might see bigger differences between implementations. It'll be as useful as the perlin noise test is today though.

 

[hatter]

^^ right now, I can't hear the audio in the clips.

In Far Cry 2, there are two screens. Which one is connected to GF100 and which is the other card?

 

[Blazkowicz]

Personally, perf/watt only matters to me in mobile products where I care about battery life. If I'm going to be plugged in and spend money on a top of the line GPU, why should I? It would be like telling a guy who buys who buys a second ultra performance sports car that he should buy first for fuel efficiency. Now, in the mid-range market it'll matter, because people won't have elaborate cooling, cases, or power supplies, but this sounds like a card aimed at the top.

Now for Tesla, it's a different story, since power density in HPC/rackmounted systems is a huge concern.

but a sports car is for those believing we still live in the 20th century.

bigger perf/watt allows either a faster card, or a less noisy, or less hot one. an exemple is the difference between GTX280 and GTX285. or hell, I'd take a quiet, slower card over a crazy loud and hot one.

in a way GF100 is helpful there. There's software voltage control, which will maybe allow to tune more. imagine a 90% clocks, 90% voltage profile (73% GPU power use) for your average game, and automatically going full blown (game profile) or o/c with your crysis 2 and stalker : edge of the rusty pipe.

 

[Sontin]

^^ right now, I can't hear the audio in the clips.

In Far Cry 2, there are two screens. Which one is connected to GF100 and which is the other card?

Left is GF100, right is ???. I guess it's a GTX285.

 

[mczak]

<80 filtering units isn't necessarily a problem if you've taken a TMU, split the fetching from the filtering, and clocked the filtering hardware at the rate of the rest of your ALUs. Absent that, I would worry about trilinear "shortcuts"

Something like 8 (per cluster) texture address / fetch operations per clock, and 4 (bilinear) filter operations per hot clock? Would be pretty much amount to the same as the old 8 fetch / 8 filter at slow clock. I wonder though what would be the incentive to use hot clock for filtering? Do this in the ALU itself, or would there be an area (what about power) advantage if still using separate logic?

 

[fellix]

The RT scene video is using three GF100 cards to do the rendering.

Looks like it is the same system from CES displaying the rocket sled demo -- the one with the big 300mm case fan and probably water cooled.

 

[neliz]

again, microsoft mouse, guess where it is..

 

[SimBy]

Left is GF100, right is ???. I guess it's a GTX285.

Seems to be running at 1920x1200 4xAA max settings.

To put things into perspective:

edit: actually some settings in that GF100 benchmark are Very High and not Ultra

 

[hatter]

Left is GF100, right is ???. I guess it's a GTX285.

Just saw the Far Cry 2 chart at TechReport and the video seems very plausible. Around 84FPS @ 1920X1200 puts it right in line with what we are hearing so far. 20 to 30% faster than 5870

 

[gamervivek]

Seems to be running at 1920x1200 4xAA max settings.

To put things into perspective:

edit: actually some settings in that GF100 benchmark are Very High and not Ultra

No it is ultra high overall, some settings don't go to ultra high.
As for the Anandtech review, it seems to be a bit off, only a 3fps loss going to 2560 from 1920.

 

[DegustatoR]