GeForce G310 - NvidiaBloggo - All the latest Nvidia news! 2010

The fact that the GeForce GF100 has high energy consumptions isn’t that much of a mystery: it’s easy to determine that by the fact that each card is connected with two PCI-Express connectors, a 6-pin one and a 8-pin one.

The Radeon HD 5800 series cards feature two 6-pin connectors, while only the dual-GPU HD 5970 has the same connectors as the GF100. It’s possible that these connectors will be different on the final version of the card, but we think it’s not likely to be that way.

The cooling system should be similar to what was seen on the single-GPU GTX 200 family: hence, a fan that brings the air from the inside and cools the radiator directly in contact with both the GPU and the GDDR5 memories, and also the power supply circuits. After the end of the cycle, the air is pushed out through the grid present above the display connectors.

nVIDIA has shown many demonstrations during the CES in Las Vegas, on systems based on GeForce Fermi cards, with single-GPU, SLI and Triple SLI configurations. All the Triple SLI configurations were based on a liquid cooling system as it can be shown on the picture above. We cannot know for sure if the liquid system was chosen due to aesthetic reasons or due to the heat generated by the cards in triple-SLI.

Starting from the card’s size: the length seems to be similar to those that can be found on the GeForce GTX 200 currently present on the market, and that shouldn’t pose any problems in the installation on most part of the cases in the market. In this sense, nVIDIA has decided to keep a certain standard instead of following ATI’s steps that brought better performing cards with their HD 5800 series, but longer and occupying more space.

The design also sees a heatsink that occupies two slots, covering the entire PCB and leaving the backside uncovered. The uncovered area has no meaningful components such as memory modules. The 384-bit bandwidth requires 12 GDDR5 memory models, each one of them connected to the memory controller with 32bit bus.

The Card

Similar to what happened back in October, when during their GPU Technology Conference, nVIDIA made the first public demonstration of a working Fermi solution within a demo system but showing a non-working dummy board to the participants, nVIDIA hasn’t shown the current GeForce prototype that will be based on the GF100, even if this model could be considered to be very similar to the card that will probably be launched from March on.

There are only a few images of the real card used, thanks to which we can make some considerations on the architecture and the choices made by nVIDIA. These pictures were taken during the CES 2010, during stand presentations and press meetings.

The possibility of enabling the CSAA 32x seems to have a very reduced impact on the speed performances when compared to the traditional MSAA 8x, and giving a better image quality. Also in this case, we’ll have to wait until we have a true GF100 card in order to properly compare the performance differences between the anti-aliasing modes currently available.

The good performances declared by nVIDIA with the CSAA 32x are combined with the reduced performance impact obtainable by going from AA 4x to AA 8x; it’s evident how the performance gap between the two modes is much more reduced, in relative terms, than what could be seen on a GT200 family. nVIDIA has also implemented some optimizations for the TMAA (Transparent Multisample Anti Aliasing), more evident in scenes that present high anti-aliasing elements.

The introduction of the GeForce GF100 on the market will also bring a new anti-aliasing mode: CSAA 32x (Coverage Sampling Antialiasing).

The CSAA technique is the same as the one introduced with the GeForce 8 series, but less heavy in terms of hardware resources. In the new, more evolved version, with 8 color samples and 24 coverage samples, it presents a possibility of correcting many of the issues with the normal anti-aliasing procedures.

The choices made in the development of the new CSAA mode have also made it so that the impact in the performance is reduced.

GPU Computing and a better image quality

One of the elements that nVIDIA has been spending time and resources on is the usage of the GPU potential in order to control other elements besides the graphics in a game. PhysX is the name of the engine currently dedicated to Physics acceleration, but not only for the latest games: the possibility of using a bigger potential besides what can be given by the CPU also implicates in new possibilities for GPU usage.

Real-time calculations of smoke, textures and movements is one example, and nVIDIA intends to keep on pushing the accelerator on these elements, highlighting the fact that thanks to PhysX, the games can reach a playability and enjoyability level that other cards cannot offer. During this year, new titles will be launched and new implementations will be available.

According to the data given by nVIDIA, the geometric performance that the new GF100 cards are capable of reaching are eight times superior to the GTX 285. It’s a very important increase, especially if we think that when going from the NV30 to the GT200, the elaboration performance for the shader was increased 150 times, while the polygonal one only 3 times.

To each SM, there’s a PolyMorph Engine, while one Raster engine is dedicated to each four SM. The real advantages on this kind of design come from the possibility of more parallel calculations, resulting in better performance. Unfortunately, this theory can only be proved when the GF100 cards are launched.

But how was the support to this technology implemented? Simply adding the tesselation to the previous GT200 architecture would have meant a huge hindrance when it comes to the performances, especially in the geometry management. It was necessary, thus, to study a new form of implementing it that would be capable of answering correctly to the new requests that come with the DirectX 11 APIs.

The development of a dedicated hardware component, consisting on a PolyMorph Engine and a Raster Engine, and its consequent implementation inside the architecture are two of the new features present on Fermi that hadn’t been explained before. The true challenge wasn’t introducing the units, but how to integrate them in parallel, so there wouldn’t be any hindrances in the performance.

The Tesselation implemented on the GF100

As we’ve mentioned before, the tesselation, together with the displacement mapping, are two essential elements in the rendering of a 3D scene that aims to be as geometrically realistic as possible. It’s also true, however, that the Tesselation also brings a downgrade in the performance. Tests have shown a sensible downgrade in the performance as soon as the Tesselation technology is activated.

nVIDIA, with the GF100 development, faced this problem with the Tesselation and its implementation: Tesselation represents an important and essential step towards game developing and applications, and as such, Fermi must be capable of fully supporting it.