Converging Simulation and AI: NVIDIA's Platform Roadmap for Accelerated Scientific Dis

Dan Ernst, senior director of supercomputing products at NVIDIA, outlines the company's platform roadmap for the convergence of simulation, AI, quantum computing, and agents. He notes that 90 of the top 100 systems are now accelerated, with Europe hosting 42 percent of them and 35 new accelerated systems deployed or announced across Europe in the past year. Publication trends show simulation, AI, and quantum methods all growing, with agentic workloads newly emerging. Examples include Juelich's Jupiter exascale system running the ICON kilometer-scale climate simulation with a full carbon cycle, 50-qubit quantum circuit simulation, and AI-based human brain connectivity mapping, plus Argonne's robotic laboratory assistant and automated quantum-computer calibration agents. Ernst introduces the Vera Rubin NVL4 platform, a four-way configuration delivering over 5 petaflops of native FP64 per rack with accelerated FP64 matrix operations. He highlights emulated-precision results from the 2024 Gordon Bell winning quantum chemistry team, which achieved around 4x speedups using reduced mantissa widths in cuBLAS while matching or beating native FP64 error, and notes production use of NVIDIA electronic-structure libraries by Samsung and TSMC. The Vera CPU is positioned for both HPC and agentic AI, offering 1.2 TB/s memory bandwidth and sustained latency under load; Los Alamos benchmarks showed large speedups on transport and multigrid codes and an agent-framework task dropping from over seven minutes to under one.