Joined
·
12,992 Posts
Ok, I'm going to start off by saying that I don't intend to start any kind of flamewar. I'm going to state my observations and come to as unbiased a conclusion as I can to help people understand what I believe to be happening.
So nVidia is going to be unveiling their GTX 480 and GTX 470 models exactly 1 week, and we're getting all kinds of "leaked" information. The GTX 470 is fairly locked down, but the GTX 480 information varies wildly from source to source.
We have reports of it having either 480 or 512 CUDA cores and a TDP ranging from 250-300W, and we're just now starting to get clock speed information.
But which numbers are correct?
The answer is its impossible to tell. Reviewers are apparently just now getting their "finalized" cards in.
nVidia has also apparently pushed the retail launch of GF100 back a week. Why would they do that?
Well, personally, I think they are still tweaking it.
I don't think anyone will deny that having the fastest GPU is extremely important to nVidia. Since G80 the size and cost of their flagship GPU has come a distant second to the raw performance.
But if the performance numbers we're seeing are true, GF100 is surprisingly close to a gpu that is nearly half it's size. This is worrisome.
What's even worse is that the 40nm node isn't exactly friendly. Even AMD is having issues, so yields aren't that great.
So now we've established two primary variables: performance, and yields.
nVidia needs to figure out the maximum performance they can get out of the chip while still yielding enough working units to sell.
The way it works is that nVidia will order X number of dies, and when they get them back they will be tested, and ranked according to their performance, since every die they get back is a little different. This process is called binning.
Now, in order to maximise performance nVidia needs to tweak other variables, the most important being the number of active CUDA cores, the voltage being applied to the chip, and the core, shader, and memory frequencies. All of these add up to a TDP, or thermal design power. The TDP is limited by both the effectiveness of the cooler as well as an artificial cap of 300W that they need to be under for marketing reasons.
For the GTX 470 the decision was easy, as the lower binned card and not the flagship nVidia is less concerned about it. Based on all of the rumours it looks like nVidia simply disabled 4 blocks, bringing the core count down to 448 and the TDP to 225W. These numbers have been fairly constant over the past few weeks so I'm going to assume this was finalized a while ago.
The GTX 480 on the other hand is a different matter. The rumours put it's performance at anywhere from 5-20% faster than the HD 5870. 5% isn't a lot, and the early release of AMD's Catalyst 10.3 has increased performance in a number of games by roughly that amount.
This potentially puts nVidia in a situation where they don't definitively have the fastest gpu! So what do they do? They delay shipping the cards to reviewers a week and go back and start tinkering with the GTX 480's final specs.
Here's a rough pseudo mathematical forumula for what they need:
[ A*Core Clock + B*Shader Clock + C*Memory Clock + D*CUDA Cores ]/ [(Actual TDP < 300) * (Actual performance > 5870) * (Yields > Yields required to launch) ]
Those familiar with mathematics will recognize this as a system of 1 equation and 4 unknowns, with three conditions that need to be met, or the answer is undefined.
But none of that really matters, what matters is that there is no single answer, so nVidia spends a week or so re-tooling the card, and the engineers periodically leak information as they are wont to do near launch.
This is why the past week or two has been filled with all kinds of numbers. The TDP is 280W, oh wait, now its 275W, hold on, now its 295W, no scratch that, its now 250W.
Also, its going to have 512 cores, no its going to have 480 cores now, wait its going to have both 512 and 480 cores!
The only thing we can conclusively say right now is that the GTX 480 will have X Cores, run at Y MHz with 1536MB of GDDR5 on a 384bit bus running at Z MHz, its TDP will be under 300W, there will be enough stock to launch, and it should be at least 5% faster than the 5870.
Anything else you hear until the NDA drops, while it might have been true at the time, can not be confirmed.
So nVidia is going to be unveiling their GTX 480 and GTX 470 models exactly 1 week, and we're getting all kinds of "leaked" information. The GTX 470 is fairly locked down, but the GTX 480 information varies wildly from source to source.
We have reports of it having either 480 or 512 CUDA cores and a TDP ranging from 250-300W, and we're just now starting to get clock speed information.
But which numbers are correct?
The answer is its impossible to tell. Reviewers are apparently just now getting their "finalized" cards in.
nVidia has also apparently pushed the retail launch of GF100 back a week. Why would they do that?
Well, personally, I think they are still tweaking it.
I don't think anyone will deny that having the fastest GPU is extremely important to nVidia. Since G80 the size and cost of their flagship GPU has come a distant second to the raw performance.
But if the performance numbers we're seeing are true, GF100 is surprisingly close to a gpu that is nearly half it's size. This is worrisome.
What's even worse is that the 40nm node isn't exactly friendly. Even AMD is having issues, so yields aren't that great.
So now we've established two primary variables: performance, and yields.
nVidia needs to figure out the maximum performance they can get out of the chip while still yielding enough working units to sell.
The way it works is that nVidia will order X number of dies, and when they get them back they will be tested, and ranked according to their performance, since every die they get back is a little different. This process is called binning.
Now, in order to maximise performance nVidia needs to tweak other variables, the most important being the number of active CUDA cores, the voltage being applied to the chip, and the core, shader, and memory frequencies. All of these add up to a TDP, or thermal design power. The TDP is limited by both the effectiveness of the cooler as well as an artificial cap of 300W that they need to be under for marketing reasons.
For the GTX 470 the decision was easy, as the lower binned card and not the flagship nVidia is less concerned about it. Based on all of the rumours it looks like nVidia simply disabled 4 blocks, bringing the core count down to 448 and the TDP to 225W. These numbers have been fairly constant over the past few weeks so I'm going to assume this was finalized a while ago.
The GTX 480 on the other hand is a different matter. The rumours put it's performance at anywhere from 5-20% faster than the HD 5870. 5% isn't a lot, and the early release of AMD's Catalyst 10.3 has increased performance in a number of games by roughly that amount.
This potentially puts nVidia in a situation where they don't definitively have the fastest gpu! So what do they do? They delay shipping the cards to reviewers a week and go back and start tinkering with the GTX 480's final specs.
Here's a rough pseudo mathematical forumula for what they need:
[ A*Core Clock + B*Shader Clock + C*Memory Clock + D*CUDA Cores ]/ [(Actual TDP < 300) * (Actual performance > 5870) * (Yields > Yields required to launch) ]
Those familiar with mathematics will recognize this as a system of 1 equation and 4 unknowns, with three conditions that need to be met, or the answer is undefined.
But none of that really matters, what matters is that there is no single answer, so nVidia spends a week or so re-tooling the card, and the engineers periodically leak information as they are wont to do near launch.
This is why the past week or two has been filled with all kinds of numbers. The TDP is 280W, oh wait, now its 275W, hold on, now its 295W, no scratch that, its now 250W.
Also, its going to have 512 cores, no its going to have 480 cores now, wait its going to have both 512 and 480 cores!
The only thing we can conclusively say right now is that the GTX 480 will have X Cores, run at Y MHz with 1536MB of GDDR5 on a 384bit bus running at Z MHz, its TDP will be under 300W, there will be enough stock to launch, and it should be at least 5% faster than the 5870.
Anything else you hear until the NDA drops, while it might have been true at the time, can not be confirmed.