NextSilicon Claims Maverick-2 Faster Than NVIDIA GPU and RISC-v CPU

Posted on Edited on
NextSilicon claims its Maverick-2 accelerator tops NVIDIA HGX B200 and Intel Sapphire Rapids in HPC and AI, delivering high efficiency with up to 192GB HBM3e.

Quick Report

NextSilicon unveiled the Maverick-2 Intelligent Compute Accelerator and claims it outperforms NVIDIA's HGX B200 and Intel's Sapphire Rapids in selected HPC and AI workloads.

According to NextSilicon, Maverick-2 is built on TSMC's 5 nm process and is available as a single-die PCIe card with 96 GB of HBM3e (300 W) or a dual-die OAM variant with 192 GB of HBM3e (600 W). The company provided internal benchmarks that show large gains in FP64 performance-per-watt and GUPS (for Real-time analytics) for graph workloads, which it attributes to a dataflow-style architecture that shifts overhead from dedicated control logic into adaptive software.

NextSilicon also announced Arbel, a RISC-V enterprise chip the company says will compete with current RISC-V designs and high-performance x86 cores. Arbel is described as having a wide instruction pipeline and large reorder buffers to improve utilization for scalar and vector workloads, though NextSilicon did not provide independent third-party benchmarks or a release timeline.

The claims are bold and will require external validation. Early figures come from NextSilicon's internal testing and a company blog post; independent reviews and third-party benchmarks will be needed to corroborate the performance and efficiency numbers.

Written using GitHub Copilot GPT-5 mini in agentic mode instructed to follow current codebase style and conventions for writing articles.

Source(s)

  • TPU
  • NextSilicon