NVIDIA Fermi: Architecture discussion

[AnarchX]

GeForce were 10-12% faster than their Tesla pedants.

btw.
This slide shows about 150GB/s BW and a SP performance of 1.5 TFLOPs SP:
http://www.computerbase.de/bildstrecke/27497/21/

Could be there a lower clock for DP operations?

 

[fbomber]

Fermi Tesla is out:


520-630 GFLOPs DP
typical consumption 190W
availability May 2010

http://www.heise.de/newsticker/meldung/SC09-Tesla-Rechenkarten-mit-Fermi-Chips-860474.html

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

May 2010? So GF100 will fight against rv970?

 

[fellix]

typical consumption 190W

So, the consumer SKUs should be well over 200W TDP?

 

[AnarchX]

May 2010? So GF100 will fight against rv970?

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

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

 

[Bouncing Zabaglione Bros.]

Fermi Tesla is out:


520-630 GFLOPs DP
typical consumption 190W
availability May 2010

So when you say "it's out" you mean, "it's scheduled to be out in six months time if everything goes as planned".

 

[neliz]

And that, dearly beloved... is what we call a pa(u)per launch.

Some edits: Heisse talks about 2x6 Pin power, but I can see 6+8 soldered to the board.

 

[Novum]

May 2010? So GF100 will fight against rv970?

It's very unlikely that AMD will have a entire new generation available in May 2010.

 

[rpg.314]

Tesla's usually come out a little after the geforce's launch, so expect consumer cards 1-2 months before May 2010. In that time frame, I don't think even northern islands will be ready.

 

[neliz]

Tesla's usually come out a little after the geforce's launch, so expect consumer cards 1-2 months before May 2010. In that time frame, I don't think even northern islands will be ready.

The GT200b Tesla came out a couple of months before the consumer GT200b boards, not after! There was no GT200 based Tesla.
All in all in sounds like AMD have a good lead on them with the RV890.

 

[Tchock]

If 190W is typical consumption...

 

[Dave Baumann]

I think the dates on these speak only to the availability of the memory they have chosen, rather than anything else.

 

[Bouncing Zabaglione Bros.]

And that, dearly beloved... is what we call a pa(u)per launch.

Some edits: Heisse talks about 2x6 Pin power, but I can see 6+8 soldered to the board.

Is this just another Nvidia spoiler for the AMD 5970 launch same as they did for 5870 launch? AMD launches products, Nvidia launches paper.

 

[neliz]

AMD responded : http://forum.beyond3d.com/showpost.php?p=1359262&postcount=4698

 

[ahu]

GeForce were 10-12% faster than their Tesla pedants.

btw.
This slide shows about 150GB/s BW and a SP performance of 1.5 TFLOPs SP:
http://www.computerbase.de/bildstrecke/27497/21/

Could be there a lower clock for DP operations?

I wouldn't trust that slide too much because
- Tesla is less than 1 TFLOPS, C1060 4GB is 936 GFLOPS
- Nehalem at 3 Ghz is way more than 100 GLOPS (I believe it's 288 GFLOPS SP)

So the slide doesn't start at zero on the Y-axis and/or is flawed otherwise.

Performance of 1.26 SP TFLOPS is disappointing no matter how you look at it. Even keeping in mind that the old Nvidia GLOPS include the underutilized MUL.

 

[digitalwanderer]

AMD responded : http://forum.beyond3d.com/showpost.php?p=1359262&postcount=4698

"AMD responded"?!? Are you high or am I?

You have to release a part before the competition can "respond" to it.

 

[rpg.314]

- Nehalem at 3 Ghz is way more than 100 GLOPS (I believe it's 288 GFLOPS SP)

Where the hell did you get those numbers from?

 

[neliz]

"AMD responded"?!? Are you high or am I?

You have to release a part before the competition can "respond" to it.

I am high (I am Dutch, therefore, I am high)

I should've written "AMD selectively leaks slides for a product that actually does launch this eon, to be more precise withing a ten day period."

 

[ahu]

Where the hell did you get those numbers from?

Brain fart I'm afraid

96 GFLOPS might be correct with SSE multiply-adds?
3 GHz x 4 cores x 4 SSE x 2 = 96 GFLOPS

 

[jaredpace]

This one board doesn't look sawed off, but the dark backplate still blocks the SLI connector.

I looked at images of previous generations teslas, and many have that backplate blocking one of the sli goldfingers. In the rackmounts, they have thier own dedicated pci-e slots, and are oriented laying on their sides, never connected in sli-configurations anyway.

 

[digitalwanderer]

I am high (I am Dutch, therefore, I am high)

I'm american, but one of the cool ones so I am too.

I should've written "AMD selectively leaks slides for a product that actually does launch this eon, to be more precise withing a ten day period."

I'm not disagreeing on timeframe/slideleaks, I'm disagreeing that the 5970 is a response to anything. It's been planned for a while.

The response to Fermi is still waiting in the wings at AMD, waiting for Fermi to actually launch before they spring it.