NVIDIA GF100 & Friends speculation

[Picao84]

ROFL... This is getting really silly xD
Now TDP gets lowered by 45W..

 

[Lonbjerg]

I bet that someone in NVIDIA is giggling like a schoolgirl, tossing out one false rumor after the other, just to keep the attention focused on the launch

 

[Silus]

In case of Fudo, who's almost part of nV PR, in general the more negative it's towards nV it is, the more likely it's to be true

That doesn't make any sense now does it ?
If Fudo was almost part of NV PR, then there would be no negatives at all for NVIDIA. Kind of like what happens in regards to AMD, in the "less than accurate" world of FUD and rumor

 

[Picao84]

I bet that someone in NVIDIA is giggling like a schoolgirl, tossing out one false rumor after the other, just to keep the attention focused on the launch

I agree 100%. Someone there must be laughing so hard right now.

 

[Alexko]

ROFL... This is getting really silly xD
Now TDP gets lowered by 45W..

Indeed. It was ultra-silly season a few weeks ago, now it's officially über-mega-silly season. We keep hearing every possible figure for pretty much every single feature. I for one am not buying any of it until I see official info.

 

[DegustatoR]

I'm really enjoing all this. Please continue 8)

 

[Picao84]

Because it is NOT 480 for the GTX480.

Do I detect something in that sentence? "for the GTX480". Maybe GTX470 is not 448, but 480?

 

[Arty]

Because it is NOT 480 for the GTX480.

Then Arun bang on 100% doesnt make sense.

 

[nyt]

My bet is on 512 for the PE and 480 for the out-of-stock edition.

 

[Ailuros]

ROFL... This is getting really silly xD
Now TDP gets lowered by 45W..

Maybe if some of us would use some very simple reasoning it would be easy to detect that some things just didn't act up. Please see the PSU recommendation pic from NVIDIA in one of my former posts and riddle me why a GPU with a recommended minimum PSU of 600W/42A would have a 295W TDP, while a GTX295 with a 289W TDP is recommended for a minimum 680W/46A PSU?

Chances are high that whoever picked up that 295W never really thought about the above and no I won't say that Semiaccurate has been wrong with its claimed 275W TDP either since I suspect that they recently lowered frequencies to the vrzone ones. Lowering by 25W in that case makes sense; 45W however means that the reported 295W TDP was bullocks all along. Plain and simple.

 

[Ailuros]

My bet is on 512 for the PE and 480 for the out-of-stock edition.

How much are you willing to bet exactly?

 

[PSU-failure]

Chances are high that whoever picked up that 295W never really thought about the above and no I won't say that Semiaccurate has been wrong with its claimed 275W TDP either since I suspect that they recently lowered frequencies to the vrzone ones. Lowering by 25W in that case makes sense; 45W however means that the reported 295W TDP was bullocks all along. Plain and simple.

It's not TDP but max board power, and even 250W is still quite bad if performance isn't ~30% higher than a 190W 5870. Refering to the past, these 250W are more than likely, though 300W are not impossible given TSMC's issues.

We don't know what idle power consumption is either, and seeing the huge difference between RV770 and Cypress (1st/2nd gen GDDR5 controllers) it's something to look at too.

 

[psolord]

It's not TDP but max board power, and even 250W is still quite bad if performance isn't ~30% higher than a 190W 5870. Refering to the past, these 250W are more than likely, though 300W are not impossible given TSMC's issues.

We don't know what idle power consumption is either, and seeing the huge difference between RV770 and Cypress (1st/2nd gen GDDR5 controllers) it's something to look at too.

Nvidia has a 1st gen GDDR5 controller as well. The one in the GT 240!

I don't really know what the problem with the idle power consumption of rv770 was. Was it the controller, the GDDR5 chips, or the chip itself?

 

[compres]

You might think so, but the differences tend to add up rather rapidly. Consider, for instance, if you have 60 physics updates per second, it only takes about 11.5 seconds for a 1/1000 error to be as large as the number in question.

For example, if your system adds energy due to numerical errors at a rate of 1/1000 every physics update, at 60 updates/sec, the energy gets doubled after only 11.5 seconds just due to errors.

These "one sided" errors are very rarely observed in practice, if ever.

PS: Forgot to mention, I agree that larger precission reduces issues. My point is that most people expect them to dissapear magically and this is not the case.

 

[KimB]

These "one sided" errors are very rarely observed in practice, if ever.

Fair enough, but the errors are not random, so you tend to get an unbalanced situation more often than you might think assuming a random walk.

This does explain, though, why the systems in games that are prone to these kinds of errors don't reliably demonstrate them: sometimes the errors cancel out, and you don't get any odd behavior.

PS: Forgot to mention, I agree that larger precission reduces issues. My point is that most people expect them to dissapear magically and this is not the case.

Given how relatively rare such errors are in current games, I'd be willing to bet that a switch to double precision would cause them to all but disappear.

 

[compres]

Fair enough, but the errors are not random, so you tend to get an unbalanced situation more often than you might think assuming a random walk.

This does explain, though, why the systems in games that are prone to these kinds of errors don't reliably demonstrate them: sometimes the errors cancel out, and you don't get any odd behavior.


Given how relatively rare such errors are in current games, I'd be willing to bet that a switch to double precision would cause them to all but disappear.

Changing the topic a bit: From what I have seen so far, these issues are mainly discussed in scientific computing. I wonder if it is taken seriously in game development.

 

[KimB]

Changing the topic a bit: From what I have seen so far, these issues are mainly discussed in scientific computing. I wonder if it is taken seriously in game development.

If you read up a little bit on physics API's, you'll see a number of bits specifically put in there to avoid numerical errors.

 

[mczak]

Those core clock frequency are correct.GF100's core clock is just GPC clock,not ROP/L2 clock.

But would be way more informative to list ROP/L2 clock in this case, I can divide by two myself just fine...

 

[CarstenS]

Oh, I love silly-season!

My bet is on 512 for the PE and 480 for the out-of-stock edition.

I ROFLed when i read that one! Especially, out-of-stock edition *ymmd*

The engraved surface of the cooler enclosure seems to be metal. Looks like this part is used to dissipate additional heat out of the unit?!

Basically yes. (That's what I'd conclude too, seeing those pictures)

Will that mean XMAN/Razor1 will be eating shorts ? (as they both seemed rather adamant that 480 rumor was FUD and the "real GTX 480" was 512 Cuda Cores) .. I vote for cherry flavored edibles lol

Please remind me: Did they say "active" ALUs? *SCNR*

 

[silent_guy]

Then Arun bang on 100% doesnt make sense.

I always thought Arun was talking about the maximum amount of shaders on the die that can be functionally active, not the amount of shaders on this or that SK, which is a function of yield and likely to change over time.