Shocking NVIDIA GT300 (Fermi) specifications
Stunning details have been revealed about the structure and capabilities of the new core.
According to Brightsideofnews, the GT300 will be a brutal computing monster, one we have never seen before. It is no coincidence that architecture was given the code name Fermi: Enrico Fermi was an Italian physicist who had unquestionable merits in the invention of the nuclear reactor. Although the GT300 will thankfully not be nuclear, it may fit into a reactor.
GPU specifications:
- 3,0 billion transistors
- 40 nm TSMC production
- 384 bit memory interface
- 512 shader cores [renamed to CUDA cores]
- 32 CUDA core shaders per block
- 1 MB L1 cache [shared 16 KB cache]
- 768 KB L2 Unified Cache
- Up to 6 GB GDDR5 memory
- Half-speed, double-precision IEEE 754
As you can see, if the above information is true, the GT300 will contain 3 billion transistors, more than double the 200 billion in the GT1,4. This, of course, will result in a huge core size that will accommodate 16 Stream Multiprocessors (this is the new name for the shades blocks), and these units will contain 32 CUDA cores per piece, for a total of 512 CUDA cores, or classic called stream processor. The new architecture, like the late groundbreaking G80, will feature 6 64-bit memory busses, so its memory bus will be 384 bits wide. This interface is paired with GDDR5 memory at 1,5, 3 or even 6 GB. Even these numbers are eye-catching, but if we imagine for a moment the GDDR5’s 4 GHz or higher clocks combined with the 384-bit interface, we’ll witness pretty rough bandwidths.
GPGPU is dead, cGPU is alive!
In light of the above, the GT300 is taking new paths, paths that GPU has not yet taken, targeting a different direction in functionality. The Fermi architecture performs 512 Fused Multiply-Add [FMA] operations per clock in plain precision mode, exactly half of that in double precision mode, or 256. Another curiosity is the knowledge of IEEE standards. In the past, NVIDIA only supported IEEE 754-1985 floating-point arithmetic, but the GT300 is already familiar with the latest release of the IEEE 754-2008 standard, so it supports all important industry standards, supposedly without any gimmicks.
Does the GPU provide native C ++ support?
The Fermi architecture provides native support for C [CUDA], C ++, DirectCompute, DirectX 11, Fortran, OpenCL, OpenGL 3.1, and OpenGL 3.2. For the first time in history, the GPU is capable of running C ++ code without major bugs and performance degradation, plus if we add Fortrant or C as well, it’s clear that NVIDIA has done a nice job.
That’s all that could be said in a nutshell, we hope NVIDIA will present its latest work in the next few weeks. There is no doubt that the GT300 and Fermi architecture is not only a new GeForce, but also a response to Intel Larrabee, a solution that is expected to be extremely efficient, as well as providing full support for key industry standards, and most importantly, it will also be an affordable and affordable product for the average home user.
Regarding the possible variations: for high-demand users, the GT300 will of course be available on GeForces, and for professional and industrial companies, the Quado and Tesla models will swell with power.
The amount of memory on the cards will vary depending on the target audience, the GeForce 380 series may include 1,5 GB, the Quadro and Tesla models may include 3 or even 6 GB of GDDR5.
The news processes unofficial information, so it is not worth treating it as a fact. In any case, Fermi fits into NVIDIA’s current philosophy, so the above data and descriptions may even be true.
