Nvidia data center CPU chief Ian Finder told Tom's Hardware Premium at SC26 that it is 'absurd' to suggest Nvidia foresaw the rise of agentic AI when it began developing its Vera CPU roughly four years ago. Finder instead attributed Vera's design to the company's long-running codesign approach between CPUs and accelerators, building on work started with its earlier Grace chip.
tomshardware.com
· 2026-11-18
At Supercomputing 2026, Nvidia said its new Vera CPU—built on proprietary Olympus cores rather than off-the-shelf Arm Neoverse cores—can see performance gains of nearly 39% on certain compute-heavy workloads when software is specifically optimized for its architecture. Nvidia noted that existing software doesn't automatically exploit Vera's expanded execution resources, including wider decode, deep out-of-order execution, and multiple vector and load/store units.
tomshardware.com
· 2026-11-17
At Supercomputing 2026, AMD promoted its Instinct MI430X accelerator's native double-precision (FP64) support, contrasting it with Nvidia's approach of emulating FP64 via techniques like Ozaki schemes on GPUs optimized for lower-precision AI math. AMD cited its own profiling data claiming that fewer than 10% of HPC workloads are dense matrix multiplications suited to such emulation, and said the method triples memory usage while poorly serving workloads like sparse linear algebra, stencils, and molecular dynamics.
tomshardware.com
· 2026-11-17