Content area
As an open-sourced instruction set and being flexible in hardware extension, RISC-V begins its pace to enter the world of high performance computing. One of the distinguished feature of processing units adopting RISC-V is its ability to add custom circuits with special purpose accelerators. As the artificial general intelligence becomes practical, AI accelerators become an indispensable part of computing devices, where RISC-V is a great fit for the CPU to glue accelerators together. A system of chip designed by Alibaba T-head is one of the early chip in the massive production adopting RISC-V CPU, where the CPU, named Xuantie-910, has a high performance design with 128-bit RISC-V vector processing units, which are designed for accelerating AI applications. OpenMC has been adapted to run on Xuantie-910. In the Monte Carlo method for reactor physics, fetching the neutron cross sections is the hotspot that takes the majority of the computational burden. The traditional point-wise cross sections are slow because of memory latency caused by accessing many nonconsecutive memory addresses. An AI model for cross section is hence proposed. With 2.2 KB of runtime size, the smallest in the published work, the data can be fetched entirely in the L1 cache during on-the-fly cross section evaluation through single memory read. The in-house AI model also covers the entire energy range, unlike only the resonance range is supported in previous work. So, the effects from memory latency is minimized. The average relative error in AI modeled U-238 elastic cross section is 0.6% from point-wise cross section. With a modified version of OpenMC on Apple M3 Max, for a VERA pin-cell problem, compared to the point-wise cross section, the adoption of AI modeled cross section reduces the total runtime by 7%, although the runtime for calculating U-238 elastic cross section causes 40% more runtime. The adoption of AI modeled U-238 elastic cross section leads to K-effective 302 pcm higher than the case of adoption of point-wise cross sections. Advantage of AI model has been verified. With AI modeled cross section, the neutron slowing down problems with pure elastic scattering on U-238 has been studied on Xuantie-910. The average relative error in 65,536 group fluxes is about 0.9% from using point-wise cross section. However, with accelerating with the 128-bit vector processing units, the performance degrades by 35%, because of the narrow 64-bit load and store interface to the vector register files. The performance with Al modeled cross section is about 1/4 of the case with point-wise cross sections. In addition, the 1,024-bit width Ara RISC-V vector processing has been used to study the cost of AI modeled cross section evaluation. Being able to access the open-sourced hardware design in SystemVerilog, cycle accurate circuit simulation is performed. Using the vector processing units, the cost is reduced to 65% of the case using scalar instructions. The 128-bit load and store interface to vector processing units is a major contributor to the speeding up. The width of the load and store interface to vector processing units should be the main optimization factor in chip design to accelerate the AI modeled cross section evaluation.