Technical details of Nvidia GF100

[Lalana]

I am here to check the technical details of Nvidia GF100. I am new to this forum, so I am not having enough idea about the rules and regulations of your forum. So please ignore my mistakes.!! One of my friend told me that you members are really very helpful. So I am expecting the same from your side. Please provide some notes that can be useful for me. Any other information related to the topic would be grateful.

[Radames]

After only a refresh of the GT200 has been the G80 architecture, and thus not given too many changes to the GPU (mainly, the number of execution units increases), there are the GF100 again to see much more news. Some similarities to the GT200 are still recognizable, but most have been revised more than just a little. The GF100 is manufactured in the modern (and probably currently still difficult to master) 40-nm process at TSMC and covers in about 3.0 to 3.2 billion transistors. An exact number we do not know, unfortunately, the two values appear but again and again.

[Ransom]

Purely on the data produces the calculation has been impressive. He has, for example to 512 Shader units, more than twice as many as in the GT200. In addition, the number of ROPs (Raster Operation Processor) rich 48 - again an increase. The Texture Mapping Units (TMUs) were reduced contrast to 64. This sounds strange at first though, because after all, already had the GT200 in the fully equipped 80 TMUs under the hood. However, the efficiency of TMUs have been massively increased, so the Texture Fill is supposedly finally risen considerably.

[Chakradev]

Similarly, the memory interface, which is on the GF100 only 384 bits, and therefore the same width as that on the G80 (GT200: 512-bit). But the use of highly clocked GDDR5 memory will again make up for this disadvantage. The size of the memory of the fastest GF100 card, is now expected to 1.5 GB and 50 percent larger than the previous GeForce-can, with the fastest Radeon HD 5000 cards.

[Galbraith]

They are extremely rude mind you, the cornerstones of the GF100 architecture that will support in the House for the first time Nvidia DirectX 11th And what else is there? They've indicated that the architecture for a high value placed on ensuring that geometric calculations can be performed among others with the seller, run noticeably faster. Because there with Nvidia sees the biggest opportunity to make games better. The image quality is also rated very highly by Nvidia, as well the possibility to use GPU computing in games. Should not forget the other GPU computing capabilities of the GF100, the nvidia has only recently advertised heavily. Finally, corresponding to regain the performance crown graphics cards and let the Radeon HD 5870 look old.

[spuff]

One of the biggest differences between the "old" GT200 (and G80) and the new GF100-chip is the allocation of internal components that Nvidia has significantly revised in order to significantly increase claimed by the efficiency. Most execution units Nvidia has placed in so-called "Graphics Processing Clusters" (GPCs), which each contain certain components, such as streaming Multiprocessors (SM) and the TMUs. In order to scale the GF100 architecture can be disabled Nvidia both the GPCs and the SM and the memory controller or do not obstruct the first place.

[Mahish]

The beginning of the rendering pipeline makes the host interface, which intercepts commands from the CPU. Then copy the "GigaThread Engine" provides the necessary data from system memory in its own frame buffer. Then the engine GigaThread the data to calculate shares in "Thread" blocks, which are then forwarded to the streaming Multiprocessors. These in turn divide the blocks into warp (32 threads), before they are distributed to the CUDA cores (ie the ALUs).

[Lawford]

Fully loaded the GF100 is composed of four GPCs, which in turn are each composed of four multi-streaming processors. Each SM can rely on 32 separate ALUs, which makes a total of 512 individual shader units. The ALUs of GF100, again as the GT200, scalar units. This may mean that each ALU per clock computing a single component (red, green, blue, or alpha value calculated). Every single SM can be on the GF100 per clock cycle now supplied with different components, depending on what is currently needed most. Each SM can handle at any time with 48 Warps. According to Nvidia offers GF100 particular, two important innovations through which the efficiency should be improved. There are, in the revised architecture for each GPC own scanning engine for the triangle setup (important for geometry calculations), the rasterization itself as well as Z-Cull (visibility checks).

[Jaabili]

At the previous architectures and Nvidia with the Radeon HD 5000 series, it sets the shader clusters (or whatever you want to call them) have a fixed unit that takes care of various tasks such as scanning the actual work and its results are then subsequently distributed to the shader clusters. According to Nvidia, this was not a big problem until now. With DirectX 11, and so many other triangles it is in this situation, but may experience a bottleneck, because the workload may apply differently. In addition, each dominated Polymorph Engine "Viewpoint Transform" and "attributes Setup. Both functions are necessary for the operation of the tessellation, as amended, for example, coordinates and efficiency to be increased. A stream-output function, the processed data then writing if necessary, in the frame buffer of the GPU, if it must be further modified. When the process is complete, the resulting primitives (points or lines sent, for example) for the grid engine. In previous GPU designs has also placed the entire scanning engine before all shader clusters, which also changes with the GF100. There, everyone has their own GPC Grid Engine in order to prevent another bottleneck.

[Taarank]

The generated pixel through the Z-Cull eventually mechanism that compares the depth of the pixels with the existing pixels in the frame buffer. If the new pixels are completely covered, Z-Cull delete it immediately, so as to minimize unnecessary computational effort. Nvidia has therefore placed great importance on the geometry performance in GF100. According to own data speed by a factor of eight compared to the GT200 could be increased. The show Californians to back this up, some self-created benchmarks against a Radeon HD 5870th These show the performance of the tessellation on the GF100, which is twice depending on the application until more than six times as fast as on the RV870. Since the results derived from Nvidia (and only show a specific calculation), they are normally to be treated with caution.