Thread: NVIDIA Fermi Pictured and Explained

Company nVidia has finally revealed its next-generation graphics processor, code-named Fermi, also known as the GF100. NVidia CEO Jen-Sun Huang (Jen-Hsun Huang) said that the architecture of Fermi is a "foundation" for the world's first computer graphics processor on which the decision will be issued series GeForce, Quadro, and Tesla.

GPU computing has reached an important turning point. GPU computing a reality. This is not just something for scientists and amateurs, it is also useful for most developers and companies that use parallel computing ", ─ said Jen-Sung. ─ "Graphics processors are used everywhere ─ in notebooks, PCs, servers and systems of cloud computing." Jen-Sung explained that the Fermi architecture was designed from scratch. "We created it entirely from scratch. This GPU 3 billion transistors", ─ stated Jen-Sung. ─ "Architecture Fermi built-in tools, libraries, and engines are a direct result of knowledge that we gained by working with thousands of CUDA developers around the world. In the coming years will be sure to Fermi opened a new branch of GPU".



Jen-Sung also commented on the "delay" response to the recent emergence of nVidia series Radeon HD 5800, the first generation of chips for graphics cards ATI with support for DirectX 11. "Nobody likes it when a competitor releases a new product. But we have a different vision, and it applies not only to the market. Of course, we do not want our fans and enthusiasts have been waiting for the next generation GPU. But we want to move the industry and will not change the way we work, even a little. We want to ensure that the product is ready before you start to sell it in the next few months. And he will ─ modular, flexible and powerful. “

Finally, Jen-Sung said that nVidia has focused on advancing computer industry and has little interest in replacing the processor. "nVidia does not focus on basic computers, but is interested in parallel computation", ─ added Jensen.

Fermi will support a number of new technologies, including:

• C + +, complementing existing support for C, Fortran, Java, Python, OpenCL and DirectCompute.

• ECC, the major requirement for datacenters and supercomputing centers, actively using graphics processors for computing.

• 512 CUDA cores with the new standard IEEE 754-2008, exceeding even the most modern processors.

• The increase in computing performance with double precision arithmetic in 8-fold compared with the previous generation of graphics processors nVidia.

• Parallel DataCache ─ this hierarchical cache GPU, which accelerates algorithms such as shortchanging physics, ray tracing, as well as matrix multiplication, where data address is not known in advance.

• GigaThread Engine with support for simultaneous execution of the nuclei, where different nuclei of the same applications can run on the GPU at the same time (for example, shortchanging the movement of fluid and rigid bodies in PhysX).

• Nexus ─ fully integrated heterogeneous computing environment for developing applications in Microsoft Visual Studio. Supports operating systems Windows Vista and Windows 7.



In terms of technical characteristics of GF100, which still is the GPU, is a sort of monster. 40 nm process technology allowed the pack to the chip area of 500mm ² 3 billion transistors. We GF100 will be 512 universal processors, combined with 32 pieces of blocks, making a total of 16 blocks, or as they called nVidia, streaming microprocessors (Streaming Microprocessors, SM). Each SM has 64KB of cache L1, and the capacity of the L2 cache is 768Kb. Fermi will have 48 ROP and 384-bit interface GDDR5 memory. Of course, supports DirectX 11 and tessellation. Frequencies of the new graphics chip will be approximately 650MGts, 1700MGts and 4800MGts respectively for the core, shader and memory. Likely, GF100 will be faster than the Radeon HD 5870, at least so says nVidia. But so far not known for certain, what is the AMD solution is contrasted with a new graphical monster nVidia.

Devices based on Fermi chips will be available no earlier than the first quarter of 2010. Such a long delay and the relatively high price are major problems nVidia.