The GT300/Fermi Thread

[Kuntz]

According to that post, GF104 is the dual part. Now, if these numbers are from nv, keep in mind that these would be "best case scenario's" i.e. the numbers you'll find in the slides and not in the reviews.

It was never said the GF104 was a dual chip.

If Nvidia had such a part, don't you think they themselves would release some benches to persuade potential ATI buyers to wait for Fermi?

They do have such a part, and have since November. More official info will arrive at CES in a week for you guys.

[WeeMaan]

Anybody see this?

http://www.guildwars2guru.com/forum/...ee-t2238.html?

Full load 55 degrees celsius?? Please..

GF100's maximum load temperature is 55 C.

[neliz]

It was never said the GF104 was a dual chip.

According to that post it is.

GF100 promises to deliver at least 40% more performance than the GTX295 for less money. GF104 promises double that.

They do have such a part, and have since November. More official info will arrive at CES in a week for you guys.

Do they differ much from the performance figures presented in October?

[Kuntz]

According to that post it is.

Ah that was a bit of weird wording. The 100 is 40% faster than the 295, the 104 is 80% faster than the 295 (double of 40%). Which means the 104 is about 29% faster than the 100. Rahja cannot confirm or deny the 104 part is a single or double at all, it's under his Business NDA still.

My guess is the 100 could be like a "GTX360" with disabled parts or lower clocks, and the 104 like a "GTX380" full bore, both single chip. We will know for sure soon enough!

I'm not sure what the October numbers were or where they came from, but it seems like they will be releasing some official information in 2~ days and/or definitely at CES.

[LiquidReactor]

I'm sure a fully clocked and uncrippled Fermi can deliver those numbers. But can Nvidia/TSMC actually deliver such a chip in mass numbers for the market?

because from the looks of things so far even A3 had to be SP crippled/underclocked just to get it to run.

Correct me if I'm wrong please, I do not want to buy another dual gpu card.

[marten_larsson]

Full load 55 degrees celsius?? Please..

Well, maybe they are watercooling them, else it is utter nonsense (or a hair dryer engine is providing the airflow).

Also GF104/Dual Fermi won't be able to double performance, SLI won't scale 100%, it never does, and secondly they won't be able to run it at full speed without exceeding 300W TDP.

[Kuntz]

Full load 55 degrees celsius?? Please..

His reply to another posted wanting more info on idle/load temps:

While I highly doubt it was in a freezer, your point is valid. If my electrical physics knowledge is any good (and it better be with a PhD....), then we should be looking at these temps:

Based on 20C ambient:

Idle range = 34-40C
Load range = 59-68C

This is based on the specs I have for a 300w heat removal solution at assumed clock speeds. Only problem is... I only helped design integration with the nMOS of GF100, not the pMOS. I'm going off the assumption that the pMOS and nMOS are joined with a Cobalt based upper gate, but I am not sure... this poses a serious problem in calculating heat production. So... these temps could be off by up to 5C, but I would think my calculations would be more in the range of ~2C of actual values.

[Kuntz]

I'm sure a fully clocked and uncrippled Fermi can deliver those numbers. But can Nvidia/TSMC actually deliver such a chip in mass numbers for the market?

because from the looks of things so far even A3 had to be SP crippled/underclocked just to get it to run.

Correct me if I'm wrong please, I do not want to buy another dual gpu card.

On yields:

Since about November and into December. I managed to grab some last minute data off the FTP in early December, but that was the last I had. That would have been A2 silicon. TSMC yields are a totally different ballpark bud, but if I had to make a good stab at it based on my knowledge, I would peg it around 25-30%. They are still low, but have improved dramatically as spins have stepped up.

Yes, there was. A1 silicon was temperamental at best, and was quite buggy. It did exist, and the demo was run on one of the few functional Fermi chips (not to be confused with GF100). Fast forward a little over a month from that point, and A2 silicon was being used for many benchmarks, debugging, etc. A3 silicon has just finished up, and is looking great. Yields are way up from what they were, and it looks like nVidia is going to have a limited availability March launch.

[LiquidReactor]

On yields:

That does not answer my question at all. Over at beyon3d someone posted some leaked stuff on a3 silicon. Basically it sucked @#! and couldnt even get 448 sp to run higher then 1200Mhz. And this wasnt Charlie this was chiphell.

[RPGWiZaRD]

I believe those figures only in case it's taken from an nvidia PR slide or something because for those slides they always pick games that's good for nvidia gpus and provide bad performance figures for a HD5970 and the load temp is prolly under ambient temp 0C or sth (who says it has to be 20C ambient?).

In a proper 3rd party review, the numbers wouldn't be as nice so I'm gonna stick to my original speculations that GTX380 single but highend gpu will be ~10% avg behind HD5970.

[Kuntz]

That does not answer my question at all. Over at beyon3d someone posted some leaked stuff on a3 silicon. Basically it sucked @#! and couldnt even get 448 sp to run higher then 1200Mhz. And this wasnt Charlie this was chiphell.

Ahh sorry it's 5am and my brain is fried!

This post should answer your question....kinda:

http://www.overclock.net/8085950-post411.html

He did mention "A3 silicon has just finished up, and is looking great." in this post:

http://www.overclock.net/8085891-post400.html

[neliz]

Thanks for the info Kuntz, now I know where the "Fermi is smaller than GT200" comes from. Also good he confirmed launch in the 2nd week of March and availability some time after that.

[annihilat0r]

That does not answer my question at all. Over at beyon3d someone posted some leaked stuff on a3 silicon. Basically it sucked @#! and couldnt even get 448 sp to run higher then 1200Mhz. And this wasnt Charlie this was chiphell.

on some thought, though, if that thing at chiphell had any chance of being true Charlie would have jumped at it

as for the "Rahja" post, remember that a House avatar might indicate that the user is trolling (just like House does in every episode)

[neliz]

on some thought, though, if that thing at chiphell had any chance of being true Charlie would have jumped at it

as for the "Rahja" post, remember that a House avatar might indicate that the user is trolling (just like House does in every episode)

He's basically saying everything opposite from what I hear from people I trust. He says the Die size is surprising, I'm saying it's gt200-esque. He says performance and yield are great, I'm hearing they'll have a hard time getting quantities for launch. And with Quantities that means Chips that actually make the desired speed.

Also, did someone here say the numbers were go at 19x12 4xSSAA?

[LiquidReactor]

He's basically saying everything opposite from what I hear from people I trust. He says the Die size is surprising, I'm saying it's gt200-esque. He says performance and yield are great, I'm hearing they'll have a hard time getting quantities for launch. And with Quantities that means Chips that actually make the desired speed.

Also, did someone here say the numbers were go at 19x12 4xSSAA?

http://www.overclock.net/8085586-post389.html

Here are a few more bits of FPS data for you. These were collected at 1920x1200 with 4x MSAA (no SSAA this time) and 16xAF. All settings are on high (perhaps not extreme/highest though, it isn't clear from the included documentation)

Crysis Warhead: 58 FPS
Left For Dead 2: 143 FPS (Judging from this result, this appears to be CPU limited at this point)
Hawx: 127 FPS
World of Warcraft: 46 FPS (this result seems off to me, I assume this is a driver issue)
Fallout 3: 72 FPS (clearly a driver acceleration issue here again)

Hmm...if Crysis was just on high thats not that huge of a jump.

on some thought, though, if that thing at chiphell had any chance of being true Charlie would have jumped at it

as for the "Rahja" post, remember that a House avatar might indicate that the user is trolling (just like House does in every episode)

What are you talking about, Charlie did jump on the whole 448 SP thing except he used Nvidias own documents to do it instead of citing some unnamed source/rumour for which he would get flamed for.

[Zed_X]

All of you will see what is Fermi in TWO days, and someone will be hardly surprised Performance is BRUTAL!

[Tim]

Man this is going to be awesome.

[cowie]

i'd would not even comment on what the Rahja had to say..its just way too fishy + vague.
in a few days i hope we get a bone or two

[Tim]

i'd would not even comment on what the Rahja had to say..its just way too fishy + vague.
in a few days i hope we get a bone or two

There have been much vaguer (is that a word) comments on previous architectures that turned out pretty darn accurate. Sounds like a pretty knowledgable guy.

[justin.kerr]

At CES there will be a Tri-SLI running Fermi PC...

[flopper]

All of you will see what is Fermi in TWO days, and someone will be hardly surprised Performance is BRUTAL!

a 5850 at 1ghz has brutal performance.
(just need better drivers)
same goes for a 285 or a 295.
Performance is good enough today for hardware purposes.

We need, good overclockability, cheap, great drivers, and so on..

Visionfinity seems as a name, ludricous, but Nvidia is fond of renaming so I guess it fits for them.

Can Fermi work out of the box, with good drivers?
or is the chip so dependent on drivers and the team to write them it will take the whole year before they are mature enough?
Is the thermal specs to much?
If its get way hot, well lot of rma is gonna be caused to happen due to heat issues.

Fermi is an adaptive chip, made for the new market, not for gaming purpose first and foremost but for tesla and the folding market. which has different requirement to the design.

It be interesting to see a working card in action, not some fake excited ceo card.
It might be also that the cards will be rare until summer in the shops.
many questions, so little answers.

[Nedjo]

as for the "Rahja" post, remember that a House avatar might indicate that the user is trolling (just like House does in every episode)

this Rahja dude is really "special":

Tesla (GT300) is very different from GF100. Do not mistake this again, or so help me Jesus I will intellectually murder you. Trust me, I am good at it.

http://www.guildwarsguru.com/forum/s...2&postcount=10

You are wrong? How many ways can I say this? Wrong? Incorrect? Mistaken? Dumb? Idiot? Invalid data detected? Negative on that Ghostrider? I mean... really!

http://www.guildwarsguru.com/forum/s...7&postcount=12

[ajaidev]

this Rahja dude is really "special":


http://www.guildwarsguru.com/forum/s...2&postcount=10



http://www.guildwarsguru.com/forum/s...7&postcount=12

lol yep he is a piece of work....

Short of the above: Stop being a flaming fanboy for a company that doesn't give 2 s about you.

This is what i found super funny i mean the great irony

[Grandpaq]

At CES there will be a Tri-SLI running Fermi PC...

This is my prediction as well as it has to beat the 5970 and a single card probably won`t do it.

[justin.kerr]

the one tri-sli setup I know of will be there, no predictions...