NVIDIA Fermi: Architecture discussion

[Silus]

Why not? What if it's 25% more power? What if they can't supply/cool that much? Or what if it has lower perf/W than planned and is 2 months late but the vendor can't give a significant discount to make up for it?

That sort of power increase doesn't seem reasonable...especially when knowing the real numbers now.
As for the other reasons, sure they may be valid ones, but that's not what the article says is the reason for the alleged "can" of the Fermi based supercomputer by OR.

 

[trinibwoy]

If by performance/watt you mean FLOPS/watt then yes, I agree. But my point is that Nvidia could very well have promised a certain level of performance for a certain level of power consumption but can't deliver. That doesn't mean Fermi is a poor choice, just not as good as it looked on paper, and that can be enough to cancel a project.

Well that's a whole different story. You're suggesting that measured performance is way below expectations. That may very well be the case but why wouldn't ORNL have this info long before they bought into the project. They certainly didn't sign up based on a powerpoint deck.

Maybe they went with ... dunno .. maybe .. their competitors?

What competitors? If they wanted CELL they would've chosen CELL, it's not exactly new on the block. And I wish people would stop quoting AMD's flop density while ignoring all the other factors that make their hardware (and software) unattractive for HPC.

 

[Silus]

Maybe they went with ... dunno .. maybe .. their competitors? Even the good old RV770 gives out 5Gflops per watt or maybe they went with CELL. Lots of variables not just the Gulftown.

Not in double precision, which is what's being discussed. Considering a rack with 4 RV770s, (244 GFlops * 4 / 150w * 4) perf/watt would be ~1.6.
Plus what Fermi offers feature wise, especially for HPC, is beyond what RV770 has, not to mention what NVIDIA has invested in CUDA and applications for its use, which AMD has yet to have.

 

[neliz]

Well that's a whole different story. You're suggesting that measured performance is way below expectations. That may very well be the case but why wouldn't ORNL have this info long before they bought into the project. They certainly didn't sign up based on a powerpoint deck.

When it was announced, the first batch of Furby chips were exactly how old? They, like mostly everyone else must've gotten "nvidia guesstimates" and "simulations" (running on 12 GT200's)

 

[trinibwoy]

When it was announced, the first batch of Furby chips were exactly how old? They, like mostly everyone else must've gotten "nvidia guesstimates" and "simulations" (running on 12 GT200's)

If that's the case then I'll have to adjust my perception of people who build supercomputers for a living. It sure doesn't sound like these guys are cowboys like you suggest though. Seems like they have a bit longer time horizon than your typical forum denizen.

ORNL also announced it will be creating the Hybrid Multicore Consortium. The goals of this consortium are to work with the developers of major scientific codes to prepare those applications to run on the next generation of supercomputers built using GPUs.

 

[3dilettante]

If that's the case then I'll have to adjust my perception of people who build supercomputers for a living. It sure doesn't sound like these guys are cowboys like you suggest though. Seems like they have a bit longer time horizon than your typical forum denizen.

How far along was the deal itself, though?
It's possible that they already had the building, wiring, and HVAC specced or already set down, and they are going to put something in to fill it with or without Nvidia.

 

[Bouncing Zabaglione Bros.]

If that's the case then I'll have to adjust my perception of people who build supercomputers for a living. It sure doesn't sound like these guys are cowboys like you suggest though. Seems like they have a bit longer time horizon than your typical forum denizen.

Nvidia probably said "These are the targets we are aiming at, if you come on board now based on our estimates, you can get the chips at cost, and you can be our loss-leader marketing into the world of HPC. Well even let you help us sell it at our launch event. "

If the chip then didn't meet the spec, was going to cost too much, needed too much power or didn't produce enough performance, was going to be too late, etc. The plug could have been pulled for any one of a number of contractual reasons if the chip spec didn't meet expectations.

Like you said, the supercomputer guys are not cowboys, and they would have had a tight contract that allowed them to cancel if contract stipulations were not met. If they cancelled, they probably had a good reason to do so.

 

[trinibwoy]

Yep, and power consumption doesn't seem like a good one.

 

[Groo The Wanderer]

Is there any official statement by either nVidia or OR about this?

I wrote them, they have not replied. Then again, they said they wouldn't talk to me any more, so I am not surprised. If they do, I will give them a chance, but since the new regime has not been honest with me once(1), I am not sure there is a point.

-Charlie

(1) Read their official statement here, the relevant stuff is in the notes and updates at the bottom.

 

[nutball]

I wrote them, they have not replied. Then again, they said they wouldn't talk to me any more, so I am not surprised. If they do, I will give them a chance, but since the new regime has not been honest with me once(1), I am not sure there is a point.

Naive question I'm sure, but what do Oak Ridge have against you?

 

[Groo The Wanderer]

If that's the case then I'll have to adjust my perception of people who build supercomputers for a living. It sure doesn't sound like these guys are cowboys like you suggest though. Seems like they have a bit longer time horizon than your typical forum denizen.

Or they believed what they were told. When was the last time NV delivered on real power claims?

-Charlie

 

[Groo The Wanderer]

Yes but Nvidia says that this 900W figure is "typical" while they say the individual card has a 190W typical power draw and 225W TDP. So either there's some kind of overhead somewhere, or they don't use those terms consistently

You are comparing a card to a 4 board enclosure. It isn't unreasonable that the enclosure uses the extra 150W or so for multiple PCIe switches, fans, blinkenlights, and other things, not to mention PSU efficiencies if read at the wall.

-Charlie

 

[Groo The Wanderer]

Well that's a whole different story. You're suggesting that measured performance is way below expectations. That may very well be the case but why wouldn't ORNL have this info long before they bought into the project. They certainly didn't sign up based on a powerpoint deck.

Think about it this way, if they had a performance target of X, and the chips came in at 20% below X, you can jack the power up for clock speed, or you can add in 20% more cards/racks. Both scenarios are power problems.

-Charlie

 

[trinibwoy]

How is that a power problem? Sticking CPUs in there instead will still be worse overall for a given performance target. Basically you're saying that they cancelled it because instead of being X times more efficient than CPUs it's only Y times more efficient. Oh noes!!!

See the problem with that theory?

 

[Arty]

What competitors? If they wanted CELL they would've chosen CELL, it's not exactly new on the block. And I wish people would stop quoting AMD's flop density while ignoring all the other factors that make their hardware (and software) unattractive for HPC.

Does OR have to use something off the bleeding edge of technology? They might have banked on Fermi coming through with x performance, y features at z timeline. Something didnt work out so they might have opted for better (or worse) alternatives.

Not in double precision, which is what's being discussed. Considering a rack with 4 RV770s, (244 GFlops * 4 / 150w * 4) perf/watt would be ~1.6.
Plus what Fermi offers feature wise, especially for HPC, is beyond what RV770 has, not to mention what NVIDIA has invested in CUDA and applications for its use, which AMD has yet to have.

How about Cypress; 3.1Gflops/W.

 

[trinibwoy]

Does OR have to use something off the bleeding edge of technology? They might have banked on Fermi coming through with x performance, y features at z timeline. Something didnt work out so they might have opted for better (or worse) alternatives.

You keep saying this but what's the better alternative you have in mind?

How about Cypress; 3.1Gflops/W.

Your original 2.9 number was correct

 

[Silus]

How about Cypress; 3.1Gflops/W.

Sure, but where is Cypress ECC support ? Where is AMD's tool for Visual Studio based development on the GPU ?

They simply do not exist and are features directly aimed at the HPC market.

 

[Creig]

You keep saying this but what's the better alternative you have in mind?

Who knows. Maybe if they couldn't reach a certain performance/watt ratio they deemed the project not efficient enough to construct.

 

[Sontin]

Who knows. Maybe if they couldn't reach a certain performance/watt ratio they deemed the project not efficient enough to construct.

And instead of using a lower amount of gpus they will cancel the project?

 

[3dilettante]

If the deal is in the early stages, they might not have much to lose.

If there were other possible outstanding offers, or perhaps they want to twist Nvidia's arm for an even better deal, it might make sense to not go forward on the current one.