GF106 and GF108 Fermi cards are due to turn up from September 13th.
When they do arrive they may work at GPUGrid straight out of the box, but there is no guarantee. They may be closer to GF100 than GF104 designs, but we won’t know all the details until very close to the release date. While the scientists might take a while to get working code, and possibly longer to get optimised code, these will support CUDA, so most should eventually work well here. Just don’t rush out to buy one expecting excellent performance before they are confirmed as working, and optimised.

The top GF106 card, the GTS450 is expected to sport specs of 1GB GDDR5 running at 3,760MHz on a 128-bit bus, supported by a quietly confident 789MHz core. Should this top GF106 card have the speculated full complement of 256 Cuda Cores then it could initially be faster than the un-optimised GTX460, depending on shader design. However in theory it would fall short of the GTX460 by about 12%, if both cards were optimised (going by cuda count and reference speeds), not that reference will mean much for GF106 and GF108. As for Compute Capability, and its effect on crunching, who knows, but that’s the point – wait and see.

It is quite possible some of these cards will offer better gaming for older games and cost a bit less than the GTX460, so they might prove to be good value for money, or indeed might not.
There will be lesser versions of these GF106 and GF108 cards, no doubt with fewer cores and/or shaders; 192, 216, and 240 are the obvious candidates for the GF106 range with 128 being more likely for the GF108 cards.

GF106 and GF108 will be mainstream chips. That means nVidia will have to make them as small and cheap as possible. The basis will obviously be the Fermi architecture, as any futher changes cost them money (development time etc.).

So what do you get when you take a GF100, remove stuff which is not really neccessary for mainstream and try to slim it down a little? First you'd take away 64 bit capability. Second you'd try to group the shaders in larger groups to reduce the amount of control logic needed to feed them, i.e. you increase the packaging density of your raw power. At the same time you might be able to increase average hardware utilization if you do it right. And third you might want to increase the number of TMUs, as these things will actually have to run games and will still be judged by their game performance.

The careful ready will probably have noticed by now that what I just described are the changes nVidia made going from GF100 to GF104. These were surprisingly large, but play out very well in the end - the chip is much more efficient at providing game performance and new features (except 64 bit). Of course nVidia could go for further archtiecture changes for the smaller chips. However, I think they already went quite far enough with the redesign leading to GF104 and its new 8 "fat" multiprocessors. I wouldn't be surprised though if they removed 64 bit capability completely for the mainstream chips. And give GF106 6 of these bad boys (288 shaders max, 240 in the next smaller version) and maybe 4 for GF108. CUDA hardware capability level would get an upgrade due to the change in feature set again.

They might prefer a smaller shader granularity for these mid range chips.. but then they've just worked it out for GF104 (these are certainly more efficient than GF100s) and software is being updated for this new design now. They shouldn't give developers too much of a headache by switching again...

MrS
____________
Scanning for our furry friends since Jan 2002

MrS, what you say is certainly more attractive; a 288 shaders GTS450 at reported frequencies would be very interesting; it could be within 5% of the GTS460 performance.

Well off topic, but liveonce’s post reminded me of something that has got to be told:
I know of several manufactures, of common household appliances, that deliberately manufacture in system failures, but with built in reset switches! So when appliances fail (overheat, or are used X number of times) the Engineer just turns up and presses a reset button on you appliance, and of course charges you £70 (Euro 100) – yes, for a Muppet to flick a switch.

Italian

I don't understand why they release their high end first, then work downwards. If software maturity is what they want, can't they wait with this instead, since hardware maturity is what most people would prefer to spend their money on? Or is this exactly what they want to avoid, when they're dealing with those who can best afford to buy new stuff?

So it's the empty the fridge, before stocking up again argument?

If software maturity is what's wanted, but many are ready to get the newest thing out there just because it's new, milk the cow, work on the software, improve on the card, release it with better software support & hardware architecture, & milk the cow again.

Glad I could be of some assistance :)
One could probably talk about this a lot more, but I'll just leave it at that for now.

MrS
____________
Scanning for our furry friends since Jan 2002

The GF106 based cards will be called GTS 450, GTS 445 and GTS 440.

http://www.techreport.com/discussions.x/19394

GF106 die shot here,
http://en.expreview.com/2010/08/04/geforce-gts-450-die-shot-tips-up/8903.html

Interesting! The rumored GF106 PCB features 6 memory chips - that's a 192 bit bus nowadays. That would probably suit a 250+ shader GF106 better than 128 bit. 128 would work, but would be borderline to really limiting performance. Especially since nVidia may not yet have fixed their GDDR5 controller, so they can't clock their memory as high as ATI can.

The die shot reveals an area of 239 mm², whereas GF104 is said to be 366 mm² large. Assuming linear scaling (which does not entirely apply here) we'd get 251 shaders in GF106. Make each shader cluster smaller by removing 64 bit capability and we may indeed be looking at 288 shaders.

MrS
____________
Scanning for our furry friends since Jan 2002

251 is also very close to 256, just in case.
As you say, the proported 789Mz clock is not very high, not for a card with such a small profile, so no magic wand was waved there.
Although I would err on the side of the 128-bit speculation, it would only be with about 2pence, in the hope I'm wrong.
Perhaps 6 was for the GTS440?

Or the PCB was meant for GF104 with 192 bit and they mistakenly attributed it to GF106. Could have been some experimental design, which didn't make it out of the door.

MrS
____________
Scanning for our furry friends since Jan 2002

New Cards have been released by NVidia today.

The main card is the GeForce GTS 450

Ryan Smith's review

192 Shaders (Stream Processors), 16 ROPs, 783MHz Core with linked shaders at 1566MHz and 1GB RAM.
As with the other Fermi’s this card is based on 40nm technology.
It is expected to retail from about $130 / 130Euro / £95.
There are of course cards that sport better speeds (925MHz).

The ratio of shaders to ROPs is 12 to 1 with this GF106 card, but they are apparently grouped into blocks of 48shaders akin to the GF104 cards.
The GF100 cards are 10/1 and the GF104 cards are 14/1. GPUZ image of the GTS450


A range of Fermi OEM/laptop/desktop cards have also been released:

GeForce GTX 470M 288 Shaders @1100MHz, GDDR5
GeForce GTX 460M 192 Shaders @1350MHz, GDDR5

GeForce GT 445M 144 Shaders @1180MHz, DDR3/GDDR5
GeForce GT 435M 96 Shaders @1300MHz, DDR3
GeForce GT 425M 96 Shaders @1120MHz, DDR3
GeForce GT 420M 96 Shaders @1000MHz, DDR3
GeForce GT 415M 48 Shaders @1000MHz, DDR3

The GeForce GT 480M with 352 Shaders @850MHz was released a while back.
With an excellent heatsink, good airflow, and the right configuration, crunching on a notebook may be possible but it’s generally not recommended.
These cards will probably turn up in Small Form Factor OEM desktop systems too. Note that these systems can also be very compact, so check the temps and air flow if you do try to use them.

If none of these cards are for you there is still good news; the entire range of existing Fermi cards are dropping in price.

So it's 192 Sahders or 4 blocks of 48 for the GF106. Too bad the 200+ rumors didn't turn out to be true - if so I had bet on 6 shader blocks rather than 5 ;)

Interesting is also the choice of 3 64 bit memory controllers. The card doesn't seem to be bandwidth starved (neither is GF104 at 192 bits), so I get the feeling the additional memory controller is more for yield than for performance. But then it's strange to see the PCB as large as the one for GTX460 and with mounting places for a full GF106. You'd think they'd only do this if they intend to launch a part which actually makes use of it. Otherwise the mainstream market should be too price sensitive to waste so much PCB space, shouldn't it?

MrS
____________
Scanning for our furry friends since Jan 2002

One of the three sets of memory controller/ROP pairs is disabled.
I expect a GTS 455 to turn up, possibly following a competitor entrance, and perhaps along with a GTX475.

One of the three sets of memory controller/ROP pairs is disabled.

Some nice pics here,
http://xtreview.com/addcomment-id-13723-view-Alternative-version-GeForce-GTS-450-list.html