Recently launched the NVDIA Tesla Pascal P100 cards, SIE launches its 2U equipment and 8 P100 GPUs.
These platforms, which for some months now, we have been supplying with GTX 1080 cards, are enhanced with the new P100 cards, which can equip 12 or 16 GB AND which equip the new CoWoS HBM2 memory with ECC error correction.
In this way, the new SIE Ladon GPU range debuts systems that include a computing capacity of more than 74 Tflops in single precision, 37 Tflops in double precision, both focused on scientific calculation and 149 Tlfops for Deep learning in medium precision.
Based on the Gigabyte G250-G52 platforms, they now equip 2 redundant sources, of 2.000 W each, which are
capable of supporting the consumption of 8 GPUs of this range.
This solution, one of the most dense on the market, is already working on several of our clients with the GTX Pascal and now allows the use of the final version of CUDA 8.
The jump is exponential, because we went from Tesla K40 cards, which gave a performance of 4,2 Tflops and 12 GB of RAM, to the P100, which allow a performance of 9,3 Tlfops and 16 GB of RAM in simple precision.
All this technology is already available for several scientific computing environments, among others:
Amber 16 and Ambertools 16
Namd 2.11 and VDM
Own codes that can be directly recompiled from CUDA 7.5 to CUDA 8
Tools like OpenACC and libraries like nvGRAPH, cuBLAS, cuFFT, cuSOLVER, etc.
Docking applications such as ACEMD: Molecular dynamics software for GPUs, where the performance improvement with Pascal technology is almost double
than with Tesla Kepler technology and much higher than compared to the GTXs in the Maxwell range.
Other companies like Allinea CFD Software, MATLAB 2016b, Mathematica, is migrating its applications to CUDA 8, now that the final version and Teslas Pascal are available.