Finally some solid info straight from the horse's mouth:Nvidia Fermi architecture overviewThis is very complex stuff so I'm only providing it for verification of the following specifications, not to suggest you all go read it.Confirmed specs:384-bit GDDR5 memory controller (not 512-bit as some had suggested)arranged as 6x 64-bit controllers512 "Cuda cores"arranged as 16 "SMs" with 32 cores each40nm process3+ billion transistorsre-organization of multiple architectural elements compared to GT200no more "TPC"no more per TPC schedulerSMs have direct access to L1 cache (shared per 3 SMs in GT200)Leaked info not yet confirmed:core clock in the 600MHz rangeshader clock in the 1700MHz rangememory clock in the 3600-4800MHz range (depending on market, consumer graphics vs. professional)48 ROPs8 z samples per cycle (same as GT200 - this was already massive overkill)8 color samples per cycle (twice GT200's capabilities)8 msaa samples per cycle (twice GT200's capabilities)I couldn't glean any info about the texturing capabilities but this hasn't been an area that's needed improvement for NV GPUs for quite some time so it's not terribly relevant if this sees any increase in capability.

Damn you max! I was just about to post the same thing. Fermi looks good if it actually delivers. Funny that Nvidia completed Larrabee before Intel could! Here is a pic of Fermi from the conference. Taken off the HardOCP forum.

So this card is actually slower than my 8800GTS? good grief.

So this card is actually slower than my 8800GTS? good grief.

Is this a serious question or are you being sarcastic?

New info:looks like this thing may have 256 TMUs. Which, if true, would be an utterly insane amount of texturing capability. Hope they're aiming for the higher end of that bandwidth spectrum. 3.6GHz GDDR5 on a 384-bit bus may not be enough to feed this beast. Even 4.8GHz may not be enough.

Techguy: No i'm dead serious.. with 600mhz + ECC that should slow things down considerably. You're getting increased latency with ECC, and in order to negate the quantum effect caused by 512 useless CUDA/SHADER cores you're taking a hit on the actual rasterops. How could any of that be good news? Just features I don't need at the moment, IMHO. And no doubt at the wrong price. I won't be in a rush..as I'm sure they'll come out with a version intended for gamers.

Techguy: No i'm dead serious.. with 600mhz + ECC that should slow things down considerably. You're getting increased latency with ECC,

ECC need not be implemented on the consumer products, it is certainly a feature intended for HPC markets. Even if they do use ECC across the product line, the latency penalty it would induce is on the order of a cycle or two (compared to non-ECC memory) and is therefore irrelevant for the following reasons: 1) for GT300 there are up to a billion of these cycles every second 2) GPUs are massively parallel and to get the most performance from them you need thousands of threads in flight to keep all the SPs busy, not low cycle access on a per thread basis

and in order to negate the quantum effect caused by 512 useless CUDA/SHADER cores

This doesn't even mean anything. Who are you trying to fool here?

you're taking a hit on the actual rasterops.

ROPs are separate from SMs, they literally have nothing to do with each other. ROPs haven't been a performance limiter for a 3d application in a long time, at least not on NV architecture. Since at least G80 NV has had multiple Z, color, and MSAA fill/sample rates per ROP. The only rasterization improvements they need to make are for triangle setup, and this is the one aspect of graphics rendering that is not parallelizable in a trivial fashion, hence why all GPUs (including Cypress and Fermi) can still only setup one triangle per cycle.

How could any of that be good news?

Because everything you just said is b.s.

Just features I don't need at the moment, IMHO. And no doubt at the wrong price. I won't be in a rush..as I'm sure they'll come out with a version intended for gamers.

Fine by me. I'm just here to pass along information and prevent the spread of FUD.

If there was no such thing as a quantum effect there would be no need for ECC now would there? I guess you guys call it "overheating".. some would call it 100degress on a silicon wafer.

If there was no such thing as a quantum effect there would be no need for ECC now would there? I guess you guys call it "overheating".. some would call it 100degress on a silicon wafer.

I can't help but laugh at your comments out of sheer disbelief. You think data correction schemes have anything to do with over-heating? Chips over-heat because of leaking transistors (see: Prescott on Intel's 90nm process), and poor thermal dissipation (see "Bumpgate" G84/G86 failure in laptops). Data errors occur because of poor programming/memory management, and failing RAM.

Uhmm I think you've confused Coding Theory with actual programming... anyways this back and fourth stops here.. you've already proved enough.

So this card is actually slower than my 8800GTS? good grief.

A true expert has spoken :( (I wonder why AVN8tr write these posts. Is he/she a troll or just another odd entertainer?)

No the trolls would be the ones without any scientific background thinking they know something because they have delusions of grandeur. Please explain what a sparse matrix is and how useful it is for flight sim... afterall you just watched the video.. the problem is you can't and that's why you can't cut through the BS in a card that supplies absolutely no benefits for gaming and isn't even targeted at gamers. You mentioned a DDR5 memory in the other thread... how is that useful for anything? Afterall it didn't seem useful when ATI had better specs. All you can do is gripe but never explain a single thing.

No the trolls would be the ones without any scientific background thinking they know something because they have delusions of grandeur. Please explain what a sparse matrix is and how useful it is for flight sim... afterall you just watched the video.. the problem is you can't and that's why you can't cut through the BS in a card that supplies absolutely no benefits for gaming and isn't even targeted at gamers. You mentioned a DDR5 memory in the other thread... how is that useful for anything? Afterall it didn't seem useful when ATI had better specs. All you can do is gripe but never explain a single thing.

I assume anyone reading my posts that cares to comment is technically competent and capable of understanding the reason why what I say is true, without requiring elaboration. I'm not going to fill in the blanks for those that don't need it. Those that are simply curious and care to ask questions I am happy to correspond with. You appear to fall into neither category and I care not to categorize you. The unveiling of Fermi was an architectural brief, not a product brief. You fail to understand the distinction and have gone off on a tangent about how features targeted at the HPC market (GPU compute) are useless to gamers. My, what a brilliant observation! /sarcasm It takes a special level of ignorance to believe that a G80/G92 derivative is faster than a Fermi derivative in ANYTHING.

Based on the card's specs as of now, would it be wise to plan on a X58 as opposed to a P55 based m/b even if not planning CF or SLI? Do you think this card could saturate the bus or or we still not there yet?

Based on the card's specs as of now, would it be wise to plan on a X58 as opposed to a P55 based m/b even if not planning CF or SLI? Do you think this card could saturate the bus or or we still not there yet?

Only time will tell for sure, but any motherboard which supports PCI-express 2.0 should be adequate as long as the individual slots are capable of operation with at least 8 lanes aka "x8". There have been several performance scaling tests performed across multiple PCI-e speeds, the most recent with the Radeon HD 5870 which you can read @ Techpowerup. The bottom line is that any single GPU card should be fine with a PCI-e 2.0 slot @ x8 but you may want full x16 for multi-GPU configurations.