Share

VxP, the extended precision processor

© CINEDIA
The VxP extended-precision processor is designed
primarily for sub-problem algebraic resolution in
physics. The processor’s high-precision operations
make it faster and more energy efficient than other
solutions. It even makes some calculations possible.
It works as a fully-transparent add-on to
conventional processors on HPC compute nodes.

In physics and molecular chemistry, algebraic operations— used massively for Big Data analysis—account for 80% of high-performance compute time. Another issue is the use of limited-precision floating-point arithmetic encoded in 32 or 64 bits, which means that as problems get larger, rounding errors increase. This can make calculations much, much longer—or even impossible—to complete. Preconditioners applied to the compute kernels to compensate for this also add complexity and increase the necessary memory resources.

Augmented-precision arithmetic can be used to reduce rounding errors, ensuring the convergence of calculations and, in all cases, reducing the number of iterations required to complete the calculations by between a factor of two and a factor of five. Until now, however, software emulators—too slow for real calculations—were the only way to take advantage of extended precision.

VxP is the first generic processor to natively support extended-precision arithmetic up to 512 bits of mantissa, with performance comparable to that of 64-bit operations (with 53 bits of mantissa) on conventional processors. The VxP is a RISC-V processor augmented with support for extended precision arithmetic and memory access. The processor is designed to take over from generic 64-bit processors in high-performance compute nodes for problems that require a higher level of precision. The handover does not require resource-intensive copying or transcoding of data because the processors share the same memory within the compute node.

 

Breakdown of problem into sub-problems distributed to several compute nodes sharing the same memory within each node to enable fine-grained cooperation between the main processor and accelerator.© CEA

 

The 2023 prototype was built on a chip manufactured using GlobalFoundries’ 22FDX® platform. This prototype version delivers an acceleration factor of up to 24 on iterative Krylov space solvers (at the state of the art) compared to a software emulation of the extended precision.

 

Development of compilation chain and low-level libraries for efficient operation with extended and variable precision on conventional data structures for the target applications (matrices/vectors).© CEA

 

Learn more

  • The technology in use:

VxP is primarily intended for computational physics, molecular chemistry, structural calculations, medical imaging, climate models, fluidics, and other scientific computing needs. It can also be used to train AIs and calculate system dynamics.

  • Patents:

“Procédé et dispositif de représentation en virgule flottante avec précision variable.” Publication/patent number: FR3093571B1. Publication date: March 19, 2021.

  • Major partnerships:

The new-generation VxP accelerator will be integrated into French startup SiPearl’s Rhea processor developed under the European Processor Initiative launched in November 2018. The initiative, which involves Europe’s main HPC stakeholders, will result in the next European exascale computing processor. Rhea will be integrated into the future Jupiter supercomputer located at Jülich in Munich.

  • Flagship publications:

  1. Hoffmann, Y. Durand, and J. Fereyre, “Accelerating spectral elements method with extended precision: a case study,” in the Proceedings of the 12th International Conference on Pure and Applied Mathematics (ICPAM 2023), 2023.
  2. E. Guthmuller et al, “Xvpfloat: RISC-V ISA Extension for Variable Extended Precision Floating Point Computation,” in IEEE Transactions on Computers, 2024, doi: 10.

à lire également