Overview
- NextSilicon reports Maverick-2 delivers 600 GFLOPS on the HPCG benchmark at 750W and claims higher FP64 performance per watt than Nvidia B200 and Intel Sapphire Rapids.
- The startup says Maverick-2 achieves roughly 10x faster PageRank than modern GPUs and posts 32.6 GUPS on database updates, which it frames as 22x over CPUs and 6x over competitor GPUs.
- Sandia National Laboratories has evaluated the technology for about three years, with a program lead calling results impressive, and Maverick-2 is running in Sandia's Vanguard-II system.
- Maverick-2 is described as in production with 96GB PCIe cards and 192GB OAM modules, and the company credits a compiler that maps C++, Fortran, Python and CUDA code to the chip for software portability.
- NextSilicon disclosed a RISC-V CPU test chip called Arbel built on TSMC 5nm with a 10-wide pipeline and four 128-bit vector units, outlined plans to vertically integrate CPUs with its accelerators, and said it has raised $300 million.