nVidia GT300 “Fermi” cGPU to power world’s most powerful supercomputer

The world of High Performance Computing is about to be shaken to the ground. We received news from Oak Ridge National Laboratory and honestly, we feel that this is probably the most important announcement for the GPU architecture and its future importance to this date.

If you had any doubt that cGPU movement is not real and that CPU is the only way to go for HPC, wake up and smell the coffee. Oak Ridge National Laboratories just announced that they will deploy nVidia Tesla cards based on Fermi architecture inside their next supercomputer system and perform a “warp drive” advancement in computing space.
The future Fermi-powered supercomputer is going to be 10 times more powerful than the currently world’s most powerful supercomputer, the infamous Jaguar [based on AMD Quad-Core Opteron at 2.3 GHz] that found its home at Oak Ridge National Laboratory. Jaguar’s peak computing power is set at 1.50 PFLOPS, i.e. 1507 TFLOPS.

Now, Oak Ridge National Laboratory plans to deploy nVidia Tesla cards based on GT300 cGPU and create a supercomputer of yet unimaginable power – the one that would make their current beast run like a scared pussycat. Even though the press release hasn’t stated what performance is targeted, the round target is set at 10+ Peta FLOPS, i.e. 10,000 TFLOPS. Used in research areas such as energy and climate change, the cGPU-powered ORNL Supercomputer will be the first supercomputer to pass 10 Peta FLOPS and probably is the writing on the wall for x86 architecture in supercomputer space.

10 years ago, x86 architecture started to endanger RISC dominance in HPC space. Proprietary CPU architectures were all at rage, but that changed for good when AMD introduced its Opteron CPU. This “x86 meets 64-bit” CPU architecture was the key the HPC space needed and starting in 2003, x86 CPUs started to dismantle proprietary architectures to the tune of today’s situation. Over 80% of world’s supercomputers are now x86 based and the writing on the wall is that cGPU will dominate.

AMD’s Fusion Render Cloud with 1000 GPUs and Oak Ridge’s Supercomputer with around 2000-2500 Tesla GT300 cGPUs will offer new levels of HPC performance.

nVidia now supports IEEE 754-2008 formats, just like ATI’s Evergreen architecture and there is no reason why Linpack would not be compiled for the GPU in a commercial/consumer way, thus enabling measurement that would enable cGPUs to be counted for Top500 ranking.

Original Author: Theo Valich