The official GT300/Fermi Thread

[Farinorco]

You talk as if it will be slow or something. MOST of the transistors come from more than doubling the shaders to 512 from 240, and that alone will bring the biggest performance in games.

GT200 had 240 SPs and 1.4b transistors.

Fermi has 512 SPs and 3.0b transistors.

Guess what? The ratio transistors to SPs on Fermi is virtually the same as GT200. What does that mean?

All the people saying "omg all those transistors are wasted on GPGPU" are freaking morons!!! The GPGPU features on Fermi are at no extra cost to gamers, and game performance is not sacrificed.

It doesn't work this way. There's a lot of stuff in a GPU apart from processing units (SPs, TUs and ROPs), a lot of logic, and usually doubling the number of processing units doesn't translate into doubling the number of transistors. For example, going from RV670 (320SPs, 666 million transistors) to RV770 (800SPs, 956 million transistors) meant an increase of x2.5 processing units (not only SPs were incresed by x2.5) and only x1.45 transistors.

If the ratio of processing units to transistors is kept, that means that lots of other stuff are being added or increased. Most of this changes, in the case of GT200 formerly, and Fermi now, are on the GPU computing ground. And a great deal of them, on GPU computing features that aren't focused on mainstream GPU computing (which could end up being useful to the mainstream consumer) but on what they call "the High Performance Computing" market (which they cover with their Tesla solutions).

Even if you use all the insults in the world, that doesn't changes things. Nothing is free in a processor, nothing gets calculated magically. There need to be transistors in there, and if NVIDIA are adding lots of HPC stuff to their chips, there is going to be a cost in transistors and in die surface regardless people being more or less moronic.

And when you feel tempted to call someone "a freakin moron", you should think twice, and not doing it in the end. You know, it's not like if it was a useful argument, or a useful attitude.

EDIT and PS: I'm not trying to say that Fermi "will suck" as a 3D rendering device, or that it's going to be exactly equal than GT200. I'm not a fortune-teller, and using logic, I suppose that part of the changes will be benefitial also to 3D rendering tasks. But what it's obvious is that NV is focusing a lot on the HPC market lately, and they are including lots of features based on that market on their chips. And that features, will have a cost in transistors and surface. Even if you want to call "freakin morons" to the people who thinks so.