NVIDIA Fermi: Architecture discussion

Payment dates for pre-order S2050 is May 31st, S2070 is July 31st, It will not be available before January 31st according to the same page. And here I was thinking that they'd launch Tesla before Geforce. (grrr sources)

 

That's new, isn't it?

Well, if those dates are real, let's hope they don't.

 

Why would you think that when Nvidia themselves said otherwise? It's not new Alexko.

http://www.nvidia.com/object/io_1258360868914.html

 

Don't be so defensive. You claimed the visual studio plugin is a deciding factor when choosing a GPU for a massivelily parallel machine running linux instances. And I tell you it is not.

As others have said the general tools and libraries is a whole different matter but again not the plugin for a windows only IDE.

I did not say the dev. platform for nVidia is irrelevant in general.

 

I might just do that.

 

I never said it was a deciding factor, but it's certainly part of the decision, since as mentioned these tools were developed after the input of people working in the area.
I also said that what NVIDIA invested on CUDA and applications for the GPGPU / HPC market is far and beyond what AMD has. Thinking otherwise is very naive to say the least, especially when trying to pitch AMD as a viable option.

 

Do we have any idea what this supercomputer would do?

 

Play Crysis?

 

Speak of the devil: http://digitimes.com/news/a20091221PR204.html

 

Heh, I'm sure that some techie working there may have that idea

Yes:

http://news.cnet.com/8301-13924_3-10364534-64.html?tag=mncol;txt

 

Interesting discover made by an italian guy at HWUpgrade...

http://www.hwupgrade.it/forum/showpost.php?p=30161869&postcount=11445

http://www.nvidia.com/docs/IO/43395/BD-04983-001_v01.pdf

Actually the Tesla cards are supposed to have only 448 SP activated, thus the lower DP GFlops numbers... But they still draw 225W...

Cards are 24.8cm long.

 

The 225W is probably an upper bound for the 6GB version. 24 GDDR5 modules probably consume a considerable bit of power themselves. It's very interesting that the first Tesla boards will have cores disabled though. Given that they're low volume parts that's a bit worrisome

 

That's worrying (and a tad riducule, actually) because JHH has already officially boasted some features, and those features won't be in the actual card...
Poor yields issues not solved even in the A3 rev.?

 

the document is dated 16. november. So it concerns only A2.

 

Well, I suppose that they work on a revision before they tape it out...
So they could have understood it even before taping... Don't know, just guessing...

 

Call me suspicious, but if you're targeting a "full" chip, then your documentation (of any date) isn't going to flaunt "castrated" numbers. The target audience of this documentation aren't going to care about which spin of metal you're currently using.

 

It all depends on the accounting. Here's a mental game for you. How many "professional" cards do they sell with the high end GPU. How many total cards do they sell with the high end GPU?

The answer is that without the non-professional market to pay for the R&D they would make at best a fraction of the profits on the high end. Its like this all over the place in computing. There are gravy markets with very high profits but in order to make those profits you generally have to have some other market with much lower profits paying the bulk of the R&D.

 

For boards all you need to know is the package layout and the basic mechanics of the top side of the package. The actual die size isn't required.

 

are there tesla cards based on the same gpu as the gtx200x what percentage of the price do the geforce gtx200x cards retail for ?

 

So Tesla boards are the salvage parts.