The “VXP” is an extended precision computing accelerator developed by CEA-List that speeds up scientific computing by a factor typically ranging from 3 to 10 or even more.
In scientific computing, precision is essential to accelerate the computing convergence towards acceptable results. Frequently, it helps the convergence of computations that would fail in lower precision. Therefore, the most suitable precision must be available according to the problem.
In software, it is quite possible to emulate the calculation with a precision beyond 64 bits, but it is very slow because the computer hardware remains in 64 bits.
It is therefore preferable to use hardware computers adapted for this. The CEA-List has therefore developed the VXP solution, a hardware accelerator that allows the precision of the calculation to be adjusted – up to 512 bits – which accelerates the calculation time.
The extended precision processor, VXP, is dedicated to solving difficult algebraic problems, due to their large size or poor conditioning.
The VXP processor supports for floating point arithmetic operations up to 512-bit mantissas. Its dynamic precision is fine grain tunable for optimal use of near processor memory. By the precision of its operations, the VXP shortens the duration of calculations and reduces their energy consumption.
The VXP seamlessly complements conventional processors in compute nodes through memory sharing and the portability of its software stack.
Another advantage is that this solution is easily integrated into the hardware architecture of data centers, on the PCIe bus in the form of an accelerator chip or on the next to come UCIe bus in the form of a chiplet. The use of VXP is also easily integrated into scientific software.
