GeForce G310 Fermi - Nvidia news! 2010

After that, we can find the specific texture cache for the 4 elaboration units present in each SM, to a total of 12 Kbytes: in this case, the value remains the same as the GT200 GPU. Outside the specific SM in which the thread is processor, we find the unified L2 Cache, three times larger than the GT200, up to 768 Kbytes, and then the GDDR5 memory controller, with bandwidth of 384bit.

Also for the memory controller, as mentioned for the number of CUDA cores, it’s possible that nVIDIA will apply some restrictions for some GF100 models; a possible scenario could be the usage of 10 memory chips on the card, with a 320bit bandwidth.

Each SM integrates 4 texture unites, for a total of 64 units present inside a GF100 GPU with 512 CUDA cores. The texture units feature a dedicated cache inside the specific SM. Completing the cache hierarchy, there’s also an L2 cache to the CPU, unified between the many SMs, with 768 Kbytes. The cache hierarchy makes it that each thread manages the integrated 64 Kbytes cache in each SM, divided in a shared block, and in a second L1 block with the sizes that can be of 16K/48K or 48K/16K.

The L1 cache dedicated to load and store operations wasn’t implemented by nVIDIA in the GT200 architecture; in the GF100, it allows the system to enhance the performances especially when it comes to physics and ray tracing. The shared cache was present but with 16 Kbytes on the GT200 solutions: in the GF100 it can be up to 16 Kbytes or 48 Kbytes, alternating with the dedicated L1, so the system can have more cache for recurring data in many threads.

It is clear that the Fermi architecture was developed in order to offer double-precision elaboration capabilities, useful for some GPU computing processes, and its efficiency is much higher than what was supplied with the GT200. Not only the number of stream processors is almost doubled, but there’s also a reduced penalty in elaborating FP64 instructions. The G80, the first nVIDIA GPU for DirectX 10, for example, lacked FP64 support completely.

For each streaming microprocessor, there’s a dedicated 64Kbytes cache which can be used as shared memory and L1 cached: the ratios are 1:3 or 3:1. The ratio is dependable on the application that is being run. The GT200, for example, integrated a 16Kbytes memory that wasn’t partitionable.

Some differences between the GF100 cards could thus be based on the integration of a lower number of SM for each GPU, leaving the CUDA core numbers untouched; disabling 2 SMs, the result is a GF100 GPU with 448 CUDA cores, a number that nVIDIA has also indicated in some internal documentation that was made public online, referring to the Tesla family Fermi solutions.

Let’s check in detail how a streaming multiprocessor is made. The central part has 32 CUDA cores: each group of 4 is associated to two load and store units (LD/ST), while 2 of these groups are associated with a Special Function Unit (SFU), for a total of 8 CUDA Cores. We can find also an instruction cache, followed by two Warp Schedulers and two Dispatch operation units, connected to the file register and capable of managing up to 32.768 32-bit entries.

GF100’s graphics architecture

After having seen the base architecture on the Fermi GF100 GPU when it comes to the GPU Computing applications, coming to life in the Tesla family series, we’ll go through the technical specifications for gaming that nVIDIA decided to pull forward with the new cards.

There are 512 CUDA Cores integrated on the GF100, CUDA Cores being the names chosen by nVIDIA to indicate their streaming multiprocessors, or SM; the approach is similar to what was seen with the previous generations of nVIDIA architectures, with unified shaders compatible with the DirectX 10 APIs, with some obvious differences. If on the G80 (GeForce 8800) and GT200 (GeForce GTX200) the cores were grouped into blocks of 8 inside one SM, the approach used on the GF100 is different, with 4 more cores for each SM: 32 in total for each SM.

Before we go on with the overview on the Fermi architecture, there’s something to be said: the Fermi cards are suffering a delay on their launch, and it’s nVIDIA itself to admit it. Currently, the GF100 GPU are being manufactured by the Taiwanese company TSMC, using 40nm technology, similarly to the Radeon HD 5000 cards. We know that the GF100 architecture is complex, hence that could be the reason for the delay.

Therefore, it is on nVIDIA’s interest to make sure that the first GeForce GF100 cards will give better performances than the Radeon HD 5000 counterparts, or the wait will be not only long, but in vain.

Another question mark remains about the complessive energy consumption for the GeForce Fermi solutions, and the eventual techniques that will be applied by nVIDIA in order to keep a low consumption, both in idle and full load: it’s an important point on a graphic card, and we could only make assumptions based on the physical structure on the first samples.

Despite all the still unknown points, we can give a good overview about the innovations on the GF100 GPU, especially when it comes to the graphics, the change in approach from nVIDIA and some of the features that come with the DirectX 11 APIs that could drastically change future games.

It’s worth it to remember that nVIDIA hasn’t yet announced what versions of GF100 cards will be presented upon launch, even though it’s predictable that they will be at least 2, considering the previous launches of new architectures by nVIDIA aimed for the top market.

Hence, it’s impossible to tell the GPU clock frequencies, the number of stream processors and memory: any considerations about the performances is simply speculation for now. As we’ll see during the article, nVIDIA has supplied some performance references, but they’re more based on the analysis of specific features on the GF100 architecture, and the only game that was tested on a Fermi architecture is known to penalize ATI Radeon cards the moment the anti-aliasing is turned on.

Introduction

AMD has released the first graphic cards based on DirectX 11 architecture on September 2009: the models were the Radeon HD 5870 and HD 5850, establishing new performance references on the 3D gaming world when it comes to single GPU systems. nVIDIA now brings forward some of the architectural features of their own DirectX 11 GPU generation, indicated with the codename Fermi.

While AMD has the DirectX 11 cards in the market for momnths, nVIDIA has continued the development of their Fermi solutions on the GPU known as GF100. The cards will be more likely launched in March, and after having anticipated some architectural details in September, nVIDIA now releases more information about the GF 100, helping users understand what will be the technical specifications for the first GeForce cards based on the GF100 GPU.

nVIDIA has brought many new things to this year’s Consumer Electronics Show. Among those, the new nVIDIA GF100, the codename that indicates the Fermi architecture that is going to be launched for the desktop market. The American manufacturer hasn’t yet revealed the complete technical specifications but has shown many working samples during the expo. We’ve spotted a video showing one of these, made by 3 nVIDIA GF100 GPUs configured in triple-SLI on a EVGA X58 motherboard and a Core i7 processor. The cards were using a liquid cooling system, as nVIDIA has declared that the final design for the heatsink on the GF100 isn’t yet finished.

The system ran the Rocket Sled tech demo, an interactive test developed by nVIDIA in order to show the capabilities of the Fermi architecture, with DX11 rendering and PhysX. In the video above, everything seems to go on well, until there’s a system crash. But that shouldn’t be a problem, at least we hope.

We’ll leave you to the video now. Enjoy!