Sistemas Informáticos Europeos has begun this month to supply its customers, SIE Ladon equipment based on new Pascal technology GPU , besides having machines that customers can remotely test. They represent an improvement with very similar prices, moving from 4.5 Tflops to 6.5 Tflops and from 5.1 Tflops to 9 Tflops, as we can see in the following table.
In addition, thanks to the location of all the cores in a unique new more powerful processor than in Kepler and Maxwell ranges, the new GTX 1080 offers both single and double precision performance, the most powerful NVDIA has offered to date. Parallel computing possibilities offered by this new technology allows a quantum leap in reducing processing times. In addition the 1080 GTX, changes GDDR5 memory technology to GDDR5X, which also increases the speed of data access. In tests conducted by our customers, like the UAB, you can see performance improvements in programs such as Amber 16, which is now available for the new CUDA 8 RC standard.
They represent an improvement with very similar prices, moving from 4.5 Tflops to 6.5 Tflops and from 5.1 Tflops to 9 Tflops, as we can see in the following table. In addition, thanks to the location of all the cores in a unique new more powerful processor than in Kepler and Maxwell ranges, the new GTX 1080 offers both single and double precision performance, the most powerful NVDIA has offered to date. Parallel computing possibilities offered by this new technology allows a quantum leap in reducing processing times. In addition the 1080 GTX, changes GDDR5 memory technology to GDDR5X, which also increases the speed of data access. In tests conducted by our customers, like the UAB, you can see performance improvements in programs such as Amber 16, which is now available for the new CUDA 8 RC standard.
We can see that the improvement over the GTX Titan X (the most powerful processor in one of the previous generation) card is about a 50% on its performance. Other tests conducted by researchers at the URV, in own codes, show the same trend. If we turn to the comparative made by the Amber 16 developers , we see that the improvement is confirmed in the same way You have all the information at: http://ambermd.org/gpus/benchmarks.htm
Compared to the Titan X standard edition, improved with a test of 23,558 atoms, that is around 50% becoming double against the K20. Also noteworthy is that the Maxwell technology was not able to execute effectively codes in double precision, while new Pascal technology does it, with a much higher loss of performance in previous generations. Furthermore, it should also be mentioned that there has been a reduction in consumption of 250 W to 150 W compared to the Tesla K40 and the GTX Titan X. This variation in consumption, means that the Ladon G SIE equipment, equipped with four GPUs, reduces 400 W / h, the cost of maintenance which for systems operating 24 hours a day, is around 41 euros monthly saving on the electricity bill. However, we must consider that dissipation of heat must be cooled with an efficiency of 1 / 1.5. That is, for every watt consumed 1.5 W cooling is needed, therefore savings can reach up to 100 euros per month, ie 1,200 euros per year. Faced with such improvement, we should wonder about the limitations: 1. Our calculations should be well parallelized, otherwise we cannot migrate spreadsheet calculation CPU to GPU. 2. We need to migrate from x86 CUDA technology. This can be coped with CUDA libraries or via standard as OpenACC, which makesdeveloping codes much easier in the Open CL standard. 3. The new cards NVDIA Pascal, only work on CUDA 8. So far, we are conducting tests with the CUDA version 8 RC (Reseale Candidate), that soon will be the final the CUDA 8 version . Therefore, only if we have programs like Amber 16, binary codes or developed by ourselves codes, we can use this new version of CUDA.