The GT300/Fermi Thread

[zalbard]

Future is DP and the Quadruple precision floating point format there is no way around that fact for CPU processing or GPU-CPU processing

Even if Nvidia doesn't deliver, AMD's already done that, just need some more 5xxx cards sold and this will be enough to stimulate creating some decently performing clients.

[Mk]

Jen-Hsun Huang Trying to figure out why my engineers couldn't fit more TMU's into our GF100 products.
January 14 at 4:26am
humm any thing you guys can predict from this?

[neliz]

Jen-Hsun Huang Trying to figure out why my engineers couldn't fit more TMU's into our GF100 products.
January 14 at 4:26am
humm any thing you guys can predict from this?

That I say (and others laugh at) that the original GF100 design has 128 TMUs. only 64 are active in the final product.

[JohnZS]

I heard 80 TMU's...
My news letter says that an exciting part is coming in Q3 2010 too
And in the land of Quadro it has 6GB of VRAM and even more exciting still...............nearly double the power of the 1.5GB Quadro Part.
Hmm I wonder what this is..
....
Yes that is right, GTX 395* (Name TBC) coming Q3 2010. To add some more detail here, the 1.5GB Quadro part has 448 pipes. (yet the 3GB part has more than double the FX5800 Ultra (assume 512))
Double this.. add some more VRAM and volla.
Voodoo5 6000 external PSU time folks...
John

[neliz]

I heard 80 TMU's...

Voodoo5 6000 external PSU time folks...
John

That's one way to avoid PCI-SIG

Q3. .imagine that they planned the product for Q1.

[annihilat0r]

That's one way to avoid PCI-SIG

Q3. .imagine that they planned the product for Q1.

A PCI-SIG employee confirmed to me that neither 5970 nor GTX 295 appear in any of the compliance lists.

[BenchZowner]

And I am the only one who thinks that that facebook account is phoney ?

[annihilat0r]

And I am the only one who thinks that that facebook account is phoney ?

I think you're the only one who thinks someone doesn't think it's phony.

[Mk]

Jen-Hsun Huang I cannot tell you too much yet, but expect the definite in NVIDIA Graphics in Q3 of this year! Stay tuned!
5 minutes ago "facebook"

[570091D]

Jen-Hsun Huang's facebook is...

[Sam_oslo]

This GPU-business is getting more and more like a dirty moneky-business. Not only on this forum, but everywhere else too.

Too many kids are focusing on funny stuff, and few seams care about the technical stuff abut a GPU any longer. Is it possible that some of these are making money based on this kind of monkey business?

[Mk]

Retail board to launch in March

Our sources are whispering that Nvidia has promised Fermi chips and Fermi boards in February. The final specification should be closed one of these days if it wasn't closed already, and the partners can start planning the final cards.


The first batch is naturally controlled by Nvidia and partners will only get to buy cards, but at later date someone might even dare to make its own design. This is naturally very unlikely due to big cost and the fact that Nvidia gets too scared that something might go wrong.

The mainstream Femi will end up in partners’ hands and they will manufacture these chips on their own, but this is still at least one quarter after the original GF100 Geforce single chip high-end card.

Nvidia partners barely have anything interesting to sell, and marketing wise they have nothing to advertise due to obvious inferiority compared to ATI’s offer. Hopefully this might change with the introduction of Fermi, but again, we expect shortages of these cards as they should be very desirable.

http://www.fudzilla.com/content/view/17357/1/

[GoldenTiger]

http://www.fudzilla.com/content/view/17357/1/

Sounds like an earlier-in-March launch *shrug*? Beats me it's just FUD after all.

[froste]

^^March retail launch and AIB getting cards in Feb is what I heard as well.

[Aerwidh]

What remains to be seen is how many cards will be out there on the shelves even after retail launch

[NapalmV5]

tags if "chernobyl" was among the first set of tags.. heres

"3 billion transisland accident" (3 mile island accident)
"toxic waste"
"radioactive waste"
"hazardous waste"
"nuclear disaster"
"armageddon"
"pr nightmare"
"fermi paradox" <-

pinch in

[junkun13]

[neliz]

And I am the only one who thinks that that facebook account is phoney ?

Lol, read my post three pages up or something.. *sigh* some people and reading...

And I thought G70/71 would be the first 512-bit GPU (turned out it was 2x256) boy.. did NV fool me there

edit: on the Jen-Hsun Huang thing.. ugh...m.. mmh.. I am
http://www.facebook.com/profile.php?...00000590746752

[Alexxx650]

[RPGWiZaRD]

Somehow that pic makes me think of an end of a catacomb that spits out toxic waste, just would need to add a pool of green slime.

[JohnZS]

*WARNING TROLL POST*
What I am curious about is if those who are criticising the Fermi about having cut DP performance are the same people who said that it does not matter that the Radeon HD 5700 series cards have no DP support?!?
*END TROLL POST*

I am a bit disappointed nVidia have cut DP performance on the consumer cards, but from my understanding this is so they can have higher clocks and more pipes for consumer cards. FYI the Tesla computational cards only have 448 pipes. Quadro has "the best of both worlds", but it will be clocked lower than the consumer part. (yet have more VRAM options)
John

[saaya]

People were stating how awesome it was going to be to buy a GF100 and have that massive DP power.
People assumed Nvidia would leave full DP on the Geforce parts.

i must have missed that...

Think about that for a sec, they can't even price the salvage part against the 5870 without losing money, if those numbers are right.

idk... they cant make money off of a 250$ gpu if they sell the final card for 500$? i find that hard to believe... they are selling 260s for below 200$... and the 260 gpus probably still cost them more than 50$ a piece at this point.

i think they can make money selling fermis even with pretty bad yields, its just that they will have a worse margin then selling 285 cards, much worse... almost 0 worse...
i dont think bad yields will keep nvidia from launching fermi, i dont think theyd delay it because of bad yields... if yields would be really bad so they cant make much money on the parts theyd do the same they do now with the 285 cards... they drip feed the market and hope for better days
so the fact that nvidia hasnt launched yet is not because yields are too bad, i think they simply arent ready yet... they probably couldnt sell a fermi card sooner than what they plan now even if they could sell it for 5000$ a piece...

Jen-Hsun Huang Trying to figure out why my engineers couldn't fit more TMU's into our GF100 products.
January 14 at 4:26am
humm any thing you guys can predict from this?

That I say (and others laugh at) that the original GF100 design has 128 TMUs. only 64 are active in the final product.

so they removed some physically or just disabled them? any idea why would they disable so many tmus? heat? no notable perf gain? lots of defects in tmus?

A PCI-SIG employee confirmed to me that neither 5970 nor GTX 295 appear in any of the compliance lists.

oh cool, nice detective work, thx for sharing

i remember the 9600XT didnt even pass the pci-sig signaling test, right? and it still worked perfectly fine and oems picked it up, even before ati solved the problem... so i dont think the 300W limitation by pci-sig is a big deal... but cooling 300W alone is quite a challange already, so either way you dont really want to go beyond 300W...

I think you're the only one who thinks someone doesn't think it's phony.

i dont think so... wouldnt be surprised if that account gets offers of homoerotic services in exchange for hardware

This GPU-business is getting more and more like a dirty moneky-business. Not only on this forum, but everywhere else too.

Too many kids are focusing on funny stuff, and few seams care about the technical stuff abut a GPU any longer. Is it possible that some of these are making money based on this kind of monkey business?

you mean like people trying to discredit each other at all costs? yeah... pretty annoying...

but tbh idk what trend your talking about, the press, enthusiasts and even normal end users have never been more informed about the actual tech underlying pc hardware as much as today... just 10 years ago most people on enthusiast forums had no clue how chips are manufactured, designed, set up, what gpu units a chip contains and what they do etc...
back then people were arguing on forums based on absolutely nothing but some generic marketing phrases by the big manufacturers, like XX calling their new chip awesome, so it must definately be better than YY companys upcoming card which they instead call a "notable speed improvement"

*WARNING TROLL POST*
What I am curious about is if those who are criticising the Fermi about having cut DP performance are the same people who said that it does not matter that the Radeon HD 5700 series cards have no DP support?!?
*END TROLL POST*

I am a bit disappointed nVidia have cut DP performance on the consumer cards, but from my understanding this is so they can have higher clocks and more pipes for consumer cards. FYI the Tesla computational cards only have 448 pipes. Quadro has "the best of both worlds", but it will be clocked lower than the consumer part. (yet have more VRAM options)
John

5700 doesnt do dp at all? didnt even know that
and havent seen anybody complain...
in the long term cutting down dp might come back and bite nvidia in their 4ss in the future when they try to push cuda some more and devs will ask them how committed they really are... but seeing how fast dp perf evolves and how slow the infra structure evolves... i dont think it will matter...

and is only dp cut down or also sp? cause the fpus are usually the hottest and most power hungry parts of a chip, so maybe they had to cut them down to get power under control?

[GoldenTiger]

/worship =p jk lol... nice find!

[Chumbucket843]

DP is pretty useless for a graphics chip and would just add heat and and die area. since ATi has multiple alu's in each shader they dont have to add much to support DP its just stuck at 5:1 ratio where nv has 2:1. gpgpu is a different story but i dont see how that matters if nobody supports double precision in opencl anyways.

[LordEC911]

What remains to be seen is how many cards will be out there on the shelves even after retail launch

Don't expect any sorta of availability after launch.
Few thousand cards is what I would expect as a best case.