480 GTX: 275 watt TDP?

[1h4x4s3x]

Btw. for the HPC market this is a HUGE drawback.
Imagine the OAK RIDGE running 2'500 cards at 3d peak 23/7.
That's 3.5GW/year or $350'000 (.1$/kWh in TN).
2'500 5870s ≈ $210'000.

 

[Dark4ng3l]

When you have a farm with 2500 cards in it 210 000$ is peanuts.

 

[Lonyo]

When you have a farm with 2500 cards in it 210 000$ is peanuts.

Double it.
That heat has to go somewhere.

 

[Genx87]

Double it.
That heat has to go somewhere.

Nvidia did a conversion with their current Tesla for I believe Sandia National Laboratory. The reduction in electricity and heat for the same performance was astronomical. In this case 2500 Tesla's would eliminate a lot of CPUs and with it electrical and heat.

 

[MrK6]

Nvidia did a conversion with their current Tesla for I believe Sandia National Laboratory. The reduction in electricity and heat for the same performance was astronomical. In this case 2500 Tesla's would eliminate a lot of CPUs and with it electrical and heat.

Link? Hasn't Sandia Nat. Labs been on NVIDIA GPUs for awhile now? Either way, it's not surprising that a GPU has performance/watt than a CPU in some types of scientific calculation (look at F@H).

 

[Lonyo]

Nvidia did a conversion with their current Tesla for I believe Sandia National Laboratory. The reduction in electricity and heat for the same performance was astronomical. In this case 2500 Tesla's would eliminate a lot of CPUs and with it electrical and heat.

Yup, compared to CPUs they might be more efficient for the tasks they are being used for, but it's likely they have financial and thermal limits.
If they budgeted $x for running costs, and suddenly they are hit by increased power and cooling costs, that's either going to mean spending more than they have, or cutting back on performance.

If they have designed a new datacentre and outfitted it with a current capacity designed with expected thermal targets in mind, then they might be limited in how many cards they can get by increased power consumption, and miss intended performance targets as well.

It's all very well saying "it's only 50w per card" or similar, but it adds up, and while the product may be lots more efficient than regular CPUs for their uses, they might not be able to easily just say "well lets eat an extra 50w per card for 2500 cards".

They will still end up with a faster setup than with CPUs, but it might not be able to be what they were hoping for/what they were expecting. 50w x 2500 isn't peanuts.

 

[tviceman]

Btw. for the HPC market this is a HUGE drawback.
Imagine the OAK RIDGE running 2'500 cards at 3d peak 23/7.
That's 3.5GW/year or $350'000 (.1$/kWh in TN).
2'500 5870s ≈ $210'000.

Not an accurate comparison. How many 5870's will it take to equal the computing output of fermi for what oakridge is using them for?

 

[Dribble]

It's been all but confirmed the TDP is 250w, not 275w.

You mean Charlie might be wrong? Say it isn't so - he has never been wrong about fermi, or the 2900XT (well he probably still says it's faster then the 8800GTX) or 1000 other things.

One month ago he said the GTX 480 was both broken and unfixable with single figure yields which should mean we don't see it at all. Obviously he's right about that too. I mean he said the same thing about the GTX 295 which wasn't supposed to be possible according to him, oh wait, don't tell me, there are lots of GTX 295's, he wasn't wrong about that as well was he?

At this rate I might have to start questioning everything he writes?

 

[MrK6]

One month ago he said the GTX 480 was both broken and unfixable with single figure yields which should mean we don't see it at all. Obviously he's right about that too.

Well, there's supposedly only a 480 shader part available, instead of the full-fledged 512 shader part, so where was he wrong?

I mean he said the same thing about the GTX 295 which wasn't supposed to be possible according to him, oh wait, don't tell me, there are lots of GTX 295's, he wasn't wrong about that as well was he?

It wasn't. NVIDIA had to wait for the 55nm refresh before the GTX295 was possible.

The main point is, Fermi has been "almost released" for six months anyway, how many revisions, tweaks, and cuts has NVIDIA gone through to put out a profitable chip? Seems like he got a large amount of relatively accurate information out when no one else could/would. No one's forcing you to read the information that's out there, if it's upsetting you, you should probably hit that big ol' red "X" in the upper corner of your screen.

 

[Dribble]

Well, there's supposedly only a 480 shader part available, instead of the full-fledged 512 shader part, so where was he wrong?

Broken and unfixable would mean no release at all, not that it has slightly less shaders.

If you want to get picky about the exact number of shaders, then quoting Charlie-who-is-never-wrong from that same article:

"the short answer is that the top bin as it stands now is about 600MHz for the half hot clock, and 1200MHz for the hot clock, and the initial top part will have 448 shaders. On top of that, the fab wafer yields are still in single digit percentages."

So he said 448 shaders not 512 or 480 anyway. Getting picky he says the top clock is only 600, current rumours say 675-700. fab yields in single digits - current rumour is < 50%.

 

[dguy6789]

Broken and unfixable could mean a lot of things. Maybe Nvidia is going to release a chip that can't compete with the Radeon 5XXX series in price and performance, nor availability and will not be able to until the next refresh. If that ends up being true, then his broken and unfixable statement was right on the money.

 

[Creig]

Good to see we are on the same page.

I'm confused. What exactly do you mean by that? Are you finally admitting that Rollo has a sickness with his completely pro-NV/anti-ATi crusade? That he is 100% biased against ATi and 100% pro NV? That if he ever has anything negative to say about NV, he does it very mildly and throws in a comparison of ATi and how many levels of magnitude more evil they are? That he has a total lack of respect and a total presence of contempt?

I'm just trying to get a clarification here so that there is no possibility of misunderstanding later.

 

[tviceman]

Broken and unfixable could mean a lot of things. Maybe Nvidia is going to release a chip that can't compete with the Radeon 5XXX series in price and performance, nor availability and will not be able to until the next refresh. If that ends up being true, then his broken and unfixable statement was right on the money.

Wow that spin made me dizzy. You should design roller coasters.

 

[OCGuy]

Broken and unfixable could mean a lot of things. Maybe Nvidia is going to release a chip that can't compete with the Radeon 5XXX series in price and performance, nor availability and will not be able to until the next refresh. If that ends up being true, then his broken and unfixable statement was right on the money.

HAHAHAHAHA.

Hold on...



HAHAHAHAH.



Wow.

 

[waffleironhead]

Broken and unfixable would mean no release at all, not that it has slightly less shaders.

If you want to get picky about the exact number of shaders, then quoting Charlie-who-is-never-wrong from that same article:

"the short answer is that the top bin as it stands now is about 600MHz for the half hot clock, and 1200MHz for the hot clock, and the initial top part will have 448 shaders. On top of that, the fab wafer yields are still in single digit percentages."

So he said 448 shaders not 512 or 480 anyway. Getting picky he says the top clock is only 600, current rumours say 675-700. fab yields in single digits - current rumour is < 50%.

You believe everything that charlie writes? WOW! D: I wondered if there was anyone out there who actually did that, unlike the rest of the people here who take everything he writes with a grain of salt.

 

[Daedalus685]

HAHAHAHAHA.
Hold on...
HAHAHAHAH.
Wow.

But isn't what dguy said exactly what Charlie said to begin with..?

He was overly dramatic and ridiculous with the "broken and unfixable" crap.. but as I remember it he was never claiming that Fermi will never be released.. only that it would be 'impossible' to manufacture at spec for profit. Thus it would likely have a poor price/performance ratio, would be exceedingly rare, would be sold as an undercut version, or simply would be unable to make a profit.

Certainly we have no idea of which of those if any is the case... but to black and white the situation as he meant all or nothing, thus is either right or wrong, seems a tad bit silly.

Perhaps I just don't understand what is so funny...

 

[dguy6789]

HAHAHAHAHA.

Hold on...



HAHAHAHAH.



Wow.

What is so humorous? I said that if a series of things end up happening(which all so far seem extremely plausible), Charlie will have been right. I suppose you believe that as long as Nvidia releases anything at all, no matter how fast, expensive, and available it is, that it will be a successful product.

But isn't what dguy said exactly what Charlie said to begin with..?

He was overly dramatic and ridiculous with the "broken and unfixable" crap.. but as I remember it he was never claiming that Fermi will never be released.. only that it would be 'impossible' to manufacture at spec for profit. Thus it would likely have a poor price/performance ratio, would be exceedingly rare, would be sold as an undercut version, or simply would be unable to make a profit.

Certainly we have no idea of which of those if any is the case... but to black and white the situation as he meant all or nothing, thus is either right or wrong, seems a tad bit silly.

Perhaps I just don't understand what is so funny...

He's just another nvidia fanboy.

 

[MrK6]

Broken and unfixable would mean no release at all, not that it has slightly less shaders.

How do you know? That's your interpretation. Charlie's "source" could have said "the 512-shader part is f***ed, it won't be released" and Charlie spun that as "broken and unfixable." Is it biased, yes, is he wrong, no. We still don't have a 512-shader part. Furthermore, that could have been the case at the time Charlie's source reported in, and maybe NVIDIA's engineers will stumble upon something that allows them.

Broken and unfixable could mean a lot of things. Maybe Nvidia is going to release a chip that can't compete with the Radeon 5XXX series in price and performance, nor availability and will not be able to until the next refresh. If that ends up being true, then his broken and unfixable statement was right on the money.

That's a very liberal interpretation, but it still might be a possibility.

If you want to get picky about the exact number of shaders, then quoting Charlie-who-is-never-wrong from that same article:

"the short answer is that the top bin as it stands now is about 600MHz for the half hot clock, and 1200MHz for the hot clock, and the initial top part will have 448 shaders. On top of that, the fab wafer yields are still in single digit percentages."

So he said 448 shaders not 512 or 480 anyway. Getting picky he says the top clock is only 600, current rumours say 675-700. fab yields in single digits - current rumour is < 50&#37;.

How do you know the engineers weren't able to squeeze out 480 shader chips (one more unit enabled) just before release? Or maybe TSMC pumped out enough chips that they could have enough stock of 480 shader chips for release? There's tons of things that probably change in the weeks leading up to a launch, never mind the months that this particular one has taken. The funny part is you fanboys hate him, yet you put him on such a high pedestal so that you can knock him down any chance you get. Kind of ironic that you hold him in such high regard. He's a reporter, that's it. People feed him information or he gleans it from his sources and then writes bombastic articles about it.

The part that seems to really dig some of you as of late is that the information is mostly correct, and you seem to feel the need to point out every little discrepancy his articles have, like he's sitting there in the engineering meetings taking notes. I know I had a good idea of how (un)competitive Fermi was going to be months ago. That's why there's 5850 Crossfire in my system right now (I was just waiting for a good deal on them). Will Fermi turn out to not suck as much as was mentioned months ago? Quite possibly, but then again they've had these months to eek out every drop of performance from the architecture.

 

[1h4x4s3x]

Not an accurate comparison. How many 5870's will it take to equal the computing output of fermi for what oakridge is using them for?

Of course it's not.
There are lots of variables missing and apart from that, ATI isn't really in the HPC market, yet.
A comparison between CPUs and GPUs would be cool.

Getting picky he says the top clock is only 600, current rumours say 675-700. fab yields in single digits - current rumour is < 50&#37;.

You got it wrong.
TSMC reported 50% for the whole 40nm process.
Cypress' yield were <20% in January for example.

 

[OCGuy]

Perhaps I just don't understand what is so funny...

If you cant see what is funny about Zoners/Charlie readers taking a statement like "broken and cant be fixed" to mean something about price/performance ratio (which we dont even know yet) or whatever the hell else you can possibly spin it in to, then I dont know what to tell you.

 

[Voo]

Not sure how many people here have any ideas about HPC, but 100k per year more or less for energy is peanuts, just imagine what those SGI Altix with hundreds of CPUs and TBs of RAM need..

 

[yh125d]

Not sure how many people here have any ideas about HPC, but 100k per year more or less for energy is peanuts, just imagine what those SGI Altix with hundreds of CPUs and TBs of RAM need..

Whether 100k is peanuts or not is inconsequential.


In HPC, performance/watt is one of the (if not THE) most important factors. And if the power usage is as high as it would seem, fermi had better have the performance to back it up or NV won't see near as large a gain in the HPC market as they have aimed for.

 

[cbn]

In HPC, performance/watt is one of the (if not THE) most important factors. And if the power usage is as high as it would seem, fermi had better have the performance to back it up or NV won't see near as large a gain in the HPC market as they have aimed for.

One of the things I learned was how much air-conditioning factors into strategy.

Turns out in some cases the actual power supply to the building can be the limiting factor. (energy cost to power hardware + energy cost to cool hardware)

 

[Voo]

Whether 100k is peanuts or not is inconsequential.

No it's not, if you spend millions of dollars 100k more or less isn't that relevant.

Heat could be a much bigger problem than some minor cost increases. But after all Nvidia has no competition from Ati in the HPC market (you need more then just the architecture, just look how much Nvidia invests into development tools for Win/Linux, profilers, debuggers,..) and there are algorithms where GPUs are much faster than CPUs, so the performance/watt advantage is on Nvidias side irrelevent if they've got a 200W or 300W TDP. Lower is better, sure, but it's not as if those Altix machines were cheap, cool and not power hungry.

 

[1h4x4s3x]

Btw. some interesting stats.
In Nvidia's financial report Teslas and Quadros are referred to as PSB (Professional Solutions Business).
PSB made a profit of $99m in 09/Q2
Tesla made a profit of $10m in 09/Q2.
In Q3, PSB's profit went down to $30m.
Unfortunately, there are no numbers for Tesla.