Nvidia and AMD clash over AI CPUs and high-precision supercomputing
At SC26, Nvidia detailed its Vera CPU and Rubin HPC GPU while AMD pushed back on emulated FP64 precision and expanded its supercomputing share to over 40 percent.
Competition in high-performance computing reached a new level at the Supercomputing conference (SC26), where Nvidia and AMD presented conflicting visions for future supercomputers, AI agents, and processor designs.
Nvidia clarifies CPU strategy and unveils Rubin HPC
Nvidia spotlighted its Vera processor, which uses custom Olympus-based Arm cores. The company showed that architecture-specific software optimization can boost Vera’s performance by up to 39% in certain workloads, such as matrix multiplication. At the event, Nvidia CPU head Ian Finder dismissed marketing claims that the company had predicted the exact rise of data-center agentic workloads when designing Vera. Instead, Finder attributed the design to product intuition and noted that memory bandwidthMemory bandwidthThe maximum rate at which data moves between a processor and its system memory. Open in glossary will play a critical role in supporting future AI agents.
Nvidia also introduced a specialized graphics processing unitGraphics Processing UnitA specialized electronic circuit designed to rapidly manipulate and alter memory to accelerate the creation of images. Open in glossary dubbed Rubin HPC. Designed for scientific simulations and agentic workflows handling massive datasets, the chip contains over 50,000 CUDACUDAA computing system that lets software use graphics chips for general calculation tasks. Open in glossary cores. That is nearly double the core count of the standard Rubin processor. Nvidia plans to release the Rubin HPC in 2028.
AMD expands footprint and challenges emulated FP64
AMD strengthened its standing in traditional supercomputing, appearing in over 40% of the latest TOP500 list, representing a 14% increase. The company highlighted Lux, a system built with the United States Department of Energy that uses EPYC processors and Instinct MI355X accelerators. Another system, Discovery, is scheduled to arrive in 2028 with EPYC chips and Instinct hardware. China’s LineShine system remained the fastest supercomputer in the world, while El Capitan held the second spot.
AMD executives argued that high-precision workloads in science cannot be fully replaced by general AI compute. AMD criticized Nvidia’s use of Ozaki FP64FP64A high-precision mathematical format that uses 64 bits to store floating-point numbers. Open in glossary emulation, claiming that emulation fits fewer than 10% of high-performance workloads and triples memory consumption during an industry-wide memory crunch. AMD emphasized that its Instinct MI430X provides native FP64 hardware support.
Competitors show next-generation hardware
Intel hardware also appeared across partner exhibits at SC26. Third-party vendors displayed running servers equipped with Intel’s next-generation Diamond Rapids processors, also known as Xeon 7, while Intel presented demos for its Xeon 6+ and Crescent Island GPUGPUA chip that does many small sums at once. First made for graphics, now the main chip for AI. Open in glossary.
How does this story land?
One tap per reader.
Quick poll
Can emulated precision adequately replace native FP64 hardware in supercomputing?
One tap. No sign-up.
Ask this story
Still have a question?
Answers use only what is in this story. They are written by AI and can be wrong.