AMD and Nvidia outline competing computing strategies at SC26
AMD and Nvidia presented competing approaches to AI and high-performance computing at SC26. AMD promoted EPYC Venice for agentic workloads, while Nvidia emphasized its Vera CPU. Nvidia CPU chief Ian Finder rejected suggestions that the company had predicted the agentic computing shift, crediting processor design expertise instead. Nvidia said architecture-specific software optimization could improve Vera performance by up to 39%, although gains beyond matrix multiplication remain uncertain.
Nvidia also announced Rubin HPC, a GPU with over 50,000 CUDA cores targeting agentic workflows and scientific simulations, with availability planned in 2028. AMD defended native FP64 support on Instinct MI430X against Nvidia’s emulation approach, arguing that high-precision computing remains essential. AMD said Ozaki schemes suit less than 10% of HPC workloads and triple memory use.
AMD hardware now features in over 40% of the new TOP500 supercomputer list. China’s LineShine retained the lead ahead of El Capitan, while growing AI demand is drawing hardware toward systems that are not necessarily tested against the benchmark. Intel’s Diamond Rapids processors appeared at partner booths, and Silicon Motion and Phison demonstrated or detailed emerging enterprise storage technology.
Infrastructure constraints were another central theme, including storage reliability and networking across geographically separated data centers. Beyond the conference, Anthropic said Claude Mythos helped adapt its software stack to AMD Instinct hardware in around three days with little human oversight, suggesting a potential challenge to Nvidia’s CUDA software advantage.