NextSilicon, a startup company founded in 2017 has detailed its Maverick-2 Intelligent Compute Accelerator (ICA), claiming it outperforms NVIDIA's HGX B200 GPU and Intel's Sapphire Rapids CPUs in high-performance computing and AI workloads. Built on TSMC's 5 nm node, Maverick-2 is available as a single-die PCIe card with 96 GB of HBM3e (300 W) or a dual-die OAM version with 192 GB of HBM3e (600 W). According to internal benchmarks, Maverick-2 delivers up to 4× higher FP64 performance-per-watt than the HGX B200 and over 20× higher efficiency than Intel's Xeon Sapphire Rapids. In GUPS benchmarks, it reached 32.6 GUPS at 460 W, reportedly 22× faster than CPUs and 6× faster than GPUs. In HPCG workloads it achieved 600 GFLOPS at 750 W, while consuming roughly half the power of competing solutions. The company attributes these gains to a data-flow based architecture that shifts overhead management from hardware to adaptive software, allowing most silicon area to be used for computation rather than control logic.

Alongside Maverick-2, NextSilicon announced Arbel, a RISC-V-based enterprise-grade chip also built on TSMC's 5 nm process. The company claims Arbel already surpasses current RISC-V designs as well as Intel Lion Cove and AMD Zen 5 cores. Arbel features a 10-wide instruction pipeline with a 480-entry reorder buffer for high core utilization, running at 2.5 GHz. The chip can execute up to 16 scalar instructions in parallel and includes four 128-bit vector units for data-parallel workloads. A 64 KB L1 cache and large shared L3 cache ensure high memory throughput and low latency to reduce bottlenecks in compute-intensive applications. No release date or full benchmarks were disclosed, but NextSilicon says Arbel represents a move toward a fully open, software-adaptive silicon platform for future HPC and AI systems.