NVIDIA's GF100 DX11 GPU

[Uber_Tiny]

While we are still waiting for information to come forth on the facts behind the Fermi, some sites that were able to see the card in action have been posting up some very interesting things.

Here is a bit of what Anandtech had to say:

There’s only so much we can say about NVIDIA’s new architecture without having the hardware on-hand for testing. NVIDIA certainly built a GPU compute monster in GF100, and based on what we now know about its graphics abilities, it looks like it’s an equally capable GPU gaming monster.

But the big question is just how much of a monster will it be, and what kind of monster price tag will it come with? Let’s make no mistake, at 3 billion transistors GF100 is going to be big, and from NVIDIA’s hints it’s probably going to be the single hottest single-GPU card we’ve seen yet. Barring any glaring flaws NVIDIA has what looks to be a solid design, but at the end of the day it almost always boils down to “how fast?” and “how much?”

NVIDIA has taken a big risk on GF100, first with its compute abilities for GPGPU use, then on its geometry abilities for gaming, and now the risk is time. Being 6 months late has hurt NVIDIA, and being 6 months late has hurt consumers through uncompetitive pricing from AMD. By no means is the situation dire, but we can quickly come up with some scenarios where it is if NVIDIA can’t convincingly beat AMD in gaming performance.

NVIDIA has shown their cards, and they’re all in. Now in the next couple of months we’ll see if they’re bluffing or if they really have what it takes to win. Stay tuned

Some other sites as well:

• AlienBabel
• AnandTech
• Benchmark Reviews
• Guru 3D
• Hard OCP
• Hardware Canucks
• Hot Hardware
• NeoSeeker
• TechPowerUp

I hope NVIDIA will be able to provide a nice pre-release card for us to bench to give some stats on.

Some more Fermi Facts:
Fermi has 16 tessellation units
1 tessellation unit per SM. So that equals 1 tessellation unit per 32 shaders: Nvidia claims up 3x-6x performance advantage. Some slides have shown this to be true. It will depend on rendering bottleneck ((Ie is tessellation the main engine bottleneck)) its very cheap to enable on GF100.
4 Triangles per clock setup engine ((4x improvement over previous engines which do it at 1x))
1.5 Gigabytes of Memory
512 Cuda Cores
384 bit Bandwith Bus
GDDR5
TMUS are decoupled and run connected to the SMs rather than the shaders. Increasing efficiency and possibility of "hot clocks"
64 TMUS
48 ROPS
Polymorph Tessellation Engine:
Vertex Fetch -> Tessellator -> Viewport Transform -> Atribute setup -> Stream Output
The tessellator units are hardware. But the entire pipeline is programmable.
Improved 8x MS performance should be on AVG about 10-15% max slower than 4x.
32x CSAA 8x MS + 24 CSAA.
Transparency AA now works with Coverage Samples. In previous generious 16x CSAA would only provide 4x color samples for Transparency AA. Now its covered by coverage samples as well.
Shader engine improved better cache. Better units. More units = Improved PhysX Support.
This architecture represents a dramatic shift for Nvidia as its moved to focusing on geometry performance. As we have started reaching diminishing returns on pixel fidelity. It is a huge architect shift.