Sneak-peek at Nvidia's Fermi Review - PAGE 1

After the hustle-bustle of the massive Consumer Electronics Show last week tempered off, Nvidia put on some information sessions showing their upcoming next generation of GPU, named Fermi. A good portion of the day was spent learning intriguing new things about this long-forthcoming GPU, and there was a few surprises too.

Probably the best place to start is to start with the architecture of the GF100 part, as it is now called. While there were working Fermi GPUs on hand for the event that myself and another ten or so tech journalists were able to see in action, there still is not a final design ready to be shown for what many folks hope will become a high-end gaming video card, expected to be called the GTX 380. In fact Nvidia for now even kept the final specifications regarding the clock speeds of their first DirectX 11 capable chip under wraps. So for those things, we'll have to wait an unknown stretch longer. However, Nvidia was happy to delve into the changes brought to the new design of the GPU. And they also assured us, once ready, the GF100 would be the fastest GPU out.

With 512 CUDA (shader) cores, 16 geometry units, 4 raster units, 64 texture units, and 48 ROP units, feed by 768 MB or more (estimated) of GDDR5 memory, the GF100 certainly would seem to have the horsepower to be able to claim top-spot in the coming GPU wars ahead. But beside the addition of these parts, the perhaps more important part of the coin is how they all fit together. Nvidia took great care to design the GF100 from the ground-up to be able to accomplish a few primary goals.

One of these goals, looking forward to the visual demands of games and software of the future, was to build a GPU with a magnitude of better potential at rendering objects with superior geometric realisim. Nvidia claims that this new upcoming generation of GPU has somewhere around 8x the GT200's capability of processing geometry. This means better, more realistic scenes; and more quicker rendering of more objects. With the tessellation features brought in part by DirectX 11, the potential is there for games to climb yet another notch higher towards life-like renderings.

The big catch may be, with PC exclusive gaming development waning, and gaming go console or multi-platform, that there might not enough titles out there that will really push this new design of GPU that way it was designed to be pushed. Like DirectX 10, will DirectX 11 have a feature set that will not be explored much, by developers? Only time will tell. But with the release of DirectX11 Dirt 2, which was hardly different than the DX9 version, one has to wonder whether many dev's will embrace the technology that only a smaller margin of their playing audience will be able to take advantage of. Time will tell. Certainly the feature-set of DirectX 11 is strong, so we hope it is pushed to its limits by at least some titles.

The GF100 has been redesigned to incorporate more geometry processing banks right into each SM. Each SM is composed of 32 CUDA cores, with its own L1 cache and access to a shared memory space. These in turn run parallel with 4 texture units, and 1 "Polymorph" engine. Never used before, this PolyMorph engine as Nvidia calls it consists of a tesselator and a vertex fetch amongst other things, which help keep things running smoothly. Four new "Raster Engines" help to properly output it all with the precise timing needed, on your display. Furthermore the L2 cache has been tripled in the GF200, compared to the GT200, in this new design direction.

All this means that the GF100 will have some serious ability to render virtual objects in levels of realism and detail, if all things go according to plan. For example, the GF 200 will have the ability to enter a 32x CSAA mode. This should be quite something to see.  

We were treated to a few software demos that showcased the new GF100's ability-set. The first was a ray-tracing demo (that will be included with all the first batch of video cards), that rendered sexy cars on 3D backdrops. These weren't just images -- the viewing angle could be changed, and from there, light rays would be totally redrawn, achieving an impressive level of realism. At first glance you'd think the scene was a photographic, not a rendering. Certainly we are just about on the doorstep of not being able to tell between real images, and drawn images, just by using our own eyes alone. While it take a while even for the GF100 to render the scenes, this can be expected for the level of detail that has to be processed in something like this.

Doesn't look that bad, does it?