Fermi based Tesla will be available in Q2 2010.

[Janooo]

Double precision performance in the range of 520GFlops - 630 GFlops

What's the 5870 value?

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

 

[thilanliyan]

Wait...what?? Isn't Tesla supposed to come out BEFORE the Geforce parts (something about them being able to sell Tesla for a lot more money?)? Or is this another version of Tesla that will come out after? Or maybe they want the Geforce users to be the beta testers?

 

[Gutcheck2009]

oh my god you have to be kidding.......

 

[Idontcare]

Editors’ note: As previously announced, the first Fermi-based consumer (GeForce&#174 products are expected to be available first quarter 2010.

Since the GPU portions of Fermi's ISA are essentially a subset of the GPGPU ISA on Fermi it makes sense timeline wise and resource wise for Nvidia to first focus on validating and shipping GPU-capable Fermi's since that will be easier to do.

This is Intel's approach between consumer versus server/xeon variants of the same architecture as well.

 

[KIAman]

About NVIDIA

NVIDIA awakened the world to the power of computer graphics when it invented the graphics processing unit (GPU) in 1999...

Wow.

 

[T2k]

What's the 5870 value?

Does not compute... what are you trying to say, seriously?

 

[WelshBloke]

What's the 5870 value?

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

I think 5870 is about 540 Gflops DP, but I could have miss remembered.

 

[SlowSpyder]

What's the 5870 value?

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

Found it, 544GigaFlops double precision.

http://www.amd.com/us/products/desk.../Pages/ati-radeon-hd-5870-specifications.aspx

 

[T2k]

Ahh, you meant power.

By the numbers Fermi does not look anything extraordinary, it's a match for cypress, no doubt about it but I expected much more a full quarter late...

 

[videogames101]

Wow.

Indeed. :biggrin:

 

[Schmide]

The killer if this is true is: ATI's 2.7 vs nVidia's new number 1.2 TeraFLOPS in single-precision.

 

[Cookie Monster]

These are theoretical peak numbers. The important thing here is that can this number be achieved in numerous applications?

 

[thilanliyan]

Wow.

Indeed. :biggrin:

It should be:
"And lo, the prophet Jen Hsun descended through the 9 circles of hell and his creation NVIDIA awakened the world to the power of computer graphics when it invented the graphics processing unit (GPU) in 1999..."

 

[Tempered81]

ati's value is 1/5th of 2.7 teraflops, so about 540gflops.

nvidia has jumped up about 8x in this category from gtx200 series. ati jumped about 2.3x up from hd4000 series

 

[scooterlibby]

At this point I wouldn't even mind if Fermi's game performance was only at or above the 5870. I am really just waiting for the day (however far away it may be) that Fermi and the 5870 have wide availability. $600 for Hemlock? Yeah, I'll wait for the effects of competition and wider availability.

 

[scooterlibby]

PS

I'm not saying an argument can't be made to justify a $600 Hemlock, but personally I wouldn't spring for it until it got to ≤ $500.

 

[Schmide]

ati's value is 1/5th of 2.7 teraflops, so about 540gflops.

nvidia has jumped up about 8x in this category from gtx200 series. ati jumped about 2.3x up from hd4000 series

From their website: Processing power (single precision): 2.72 TeraFLOPS

 

[MrK6]

Granted there's supply issues now, but AMD's 40nm chips have a ton of life left in them (1000MHz+ is easy with a little voltage). I don't understand why NVIDIA dropped the ball this time and is so late to market. Between this and the supply issues, the entire market has pretty much stagnated.

 

[Tempered81]

From their website: Processing power (single precision): 2.72 TeraFLOPS

yeah that's single precision, did you see the OP?

edit:

 

[SlowSpyder]

ati's value is 1/5th of 2.7 teraflops, so about 540gflops.

nvidia has jumped up about 8x in this category from gtx200 series. ati jumped about 2.3x up from hd4000 series

I'm not sure if this means Nvidia did a really good job increasing DP performance, or if it means that their older cards just did not have very good DP performance in the first place..? I know AMD's numbers are unrealistic, just peak numbers that will never occur. But even if Nvidia's numbers are more realistic, it seems odd that the GT2x0 cards would have such low DP performance since they were supposed to be very much geared towards GPGPU, which as I understand it is what DP performance is important for.

 

[Schmide]

I say.

The killer if this is true is: ATI's 2.7 vs nVidia's new number 1.2 TeraFLOPS in single-precision.

You say

ati's value is 1/5th of 2.7 teraflops, so about 540gflops.

I say

From their website: Processing power (single precision): 2.72 TeraFLOPS

You say

yeah that's single precision, did you see the OP?

Both numbers are right. I understand the double precision number. I was making a point on single-precision which is very useful especially in calculating iterative series. You seemed to invalidate my comment.

 

[Forumpanda]

I'm not sure if this means Nvidia did a really good job increasing DP performance, or if it means that their older cards just did not have very good DP performance in the first place..? I know AMD's numbers are unrealistic, just peak numbers that will never occur. But even if Nvidia's numbers are more realistic, it seems odd that the GT2x0 cards would have such low DP performance since they were supposed to be very much geared towards GPGPU, which as I understand it is what DP performance is important for.

Just to clarify my understanding, I think I read a post from a guy trying to work out the average utilization, and I think he came to the conclusion that for SP ATI was at 75-80% and nV was at 90-95%. (obviously depends on the code).
For DP I would expect that they will sit pretty even in utilization.

So with a few buckets of salt I think those numbers are fairly realistic looking, or I could be an idiot, take your pick.

 

[Janooo]

Just to clarify my understanding, I think I read a post from a guy trying to work out the average utilization, and I think he came to the conclusion that for SP ATI was at 75-80% and nV was at 90-95%. (obviously depends on the code).
For DP I would expect that they will sit pretty even in utilization.

So with a few buckets of salt I think those numbers are fairly realistic looking, or I could be an idiot, take your pick.

That's how it seems. ATI is much better in SP and equal in DP.

 

[Idontcare]

Its too bad so few applications (outside of gaming) are available to us consumers to take advantage of all that SP (and DP) crunching capability inside that Cypress chip.

With AMD's fusion coming within the next 12 months I see this going two ways...either having access to the vastly superior SP and DP processing capability (scaled down accordingly to the SP counts on Llano of course) is just to result in products that are going to trounce Tesla's market penetration efforts OR we will see a product fielded by AMD which has faptastic hardware specs in powerpoint slides but has zilch for support in the software land (i.e. the story of AMD's GPGPU to date and 3DNow! from a decade ago) so it won't make much of an impact anyways.

I for one am praying to gods the world over that AMD has it pulled together a little better with their planned Fusion products and they will have software launching in parallel to the hardware release to take advantage of their revolutionary APU's from day one.

The performance capability is there, we see it in Cypress today, but if nothing is compiled to extract that performance then it will just be more idle xtors adding to our electric bill.

I can't help but think that despite all of Jensen's "the cpu is dead" rhetoric that he is just a little bit concerned with the prospects of what Llano and future APU products could bring to the fledgling GPGPU marketspace.

 

[scooterlibby]

These slides are pretty badass:

http://www.donanimhaber.com/Ve_huzurlarinizda_ATi_Radeon_HD_5970-16592.htm

My question is, if the 5970 is designed for 'massive overclocking,' as the slides state (with ceramic caps, digital voltage regulators blah blah), can a 6+8 pin actually handle that? I thought the main reason the 5970 was downclocked was because of TDP limits.

My apologies if I am throwing these terms around carelessly, I have only a layman's understanding of it.