Nvidia GPU Technology Conference The Rise of Fermi
Reviewed by: ccokeman
Reviewed on: September 30, 2009
Today at Nvidia's GPU Technology Conference in San Jose, California the company released details on its upcoming architecture codenamed "Fermi." This GPU is a step away from the current GT200 core and is designed from the ground up as a computational engine. That does not mean that it won't perform as a graphics core in any way. No, what this means is that the parallel computing power of Nvidia's GPUs is being used more and more for tasks that are closely associated with supercomputers. The ability to reduce the time it takes to complete computational tasks as well as reduce the power and cost footprints by using the GPU's power make their use that much more attractive. Let me throw out a number, three billion, that's the number of transistors this new architecture will carry. This architecture will carry up to 512 Cuda cores organized into 16 Streaming Multiprocessors with 32 cores each, a 4x improvement over the current generation hardware. The GPU features a 384-bit memory interface divided up into six 64-bit partitions supporting up to 6GB of GDDR5 memory. Each Streaming Multiprocessor will carry 64KB of shared configurable L1 cache and 768KB of L3 cache.
The specification sheet below shows the specifications of Fermi as it compares to the last two generations of Nvidia's graphics processors.
Currently, there has not been any mention of clock speeds but after looking at the specs this card will be a monster on specs alone. Whether that holds true is up for speculation until we see Fermi live and in person. No claims have been made as to the gaming performance delivered but during the keynote speech today Jen-Hsun Huang showed that the Fermi architecture will be 8x faster in double precision performance over the GT200. The keynote speech was the highlight of the day as Jen-Hsun Huang spoke about the current and future applications for GPU computing and how it will change the supercomputing landscape. Jen-Hsen Haung brought forth some of the current companies that are using GPU computing and described the impact of GPU computing on their application. The Oakbridge National Laboratory, Adobe, Techniscan and RTT were represented. Techniscan uses GPU computing to help with the detection of breast tissue tumors through it scanning technology. The power of the GPU reduces the time to view the results from over an hour on a four CPU cluster to less than thirty minutes with two Tesla GPUs. RTT uses the Tesla GPU to render a finished vehicle and its parts in intimate detail. Notice the picture below with the tire and how the light plays on the rendered rim (the picture in picture is the baseline). iRay uses cloud computing to render complex images with ray tracing in seconds instead of hours. Notice the shadowing and how the light interacts with the rest of the picture. All pretty amazing stuff.
The conference is for the most part dedicated to the individuals who are writing the coding for GPU computing and the firsts that Fermi brings along to help the cause with ECC memory, the use of the new floating point standard IEEE 754-2008, a new cached memory hierarchy and C++. With all the GPU computing talk, the gamer is not lost with Fermi with DirectX 11 support, advanced PhysX effects, ray tracing and 3DStereo Vision. Kind of takes the big selling point of DX11 out of the red team’s arsenal. I’ll leave you with a few pictures of the Fermi / GT300 graphics card in its Tesla dress but have no fear as Fermi variants will be in the Tesla, Quadro and GeForce series. Have a gander at the latest from Nvidia, Fermi!
In a press conference after the keynote speech it was acknowledged that there is silicon "in house" and the Fermi architecture should be available in the next few months! I can't wait!