VMXDOTP: A RISC-V Vector ISA Extension for Efficient Microscaling (MX) Format Acceleration

Mar 5, 2026 by Max Wipfli, Gamze .Islamouglu, Navaneeth Kunhi Purayil, Angelo Garofalo, Luca Benini

We designed VMXDOTP, a compact RVV 1.0 extension that makes block-floating-point microscaling formats (MXFP8/MXFP4) run like native vector dot products, letting shared-L1 VPE clusters keep pipelines full and scale precision and block sizes in software. Implemented in 12 nm FinFET, VMXDOTP nets near-peak utilization and big wins in throughput, area, and energy versus emulation and prior MX engines—so you get flexible transformer-friendly acceleration without the typical MX mess.

source S2