The Fursoy Slice‑1 core integrates a 2‑wide superscalar pipeline, 128‑entry ROB and 16 KiB L1 caches, all meeting a 500 MHz target in 45 nm.
*An open‑source 64‑bit out‑of‑order RISC‑V processor, built in SystemVerilog, achieves 500 MHz timing in Nangate 45 nm and runs mainline Linux 6.6.50. The milestone challenges the notion that high‑performance silicon requires proprietary IP.*
An open‑source team has delivered a 64‑bit out‑of‑order RISC‑V processor that not only boots mainline Linux 6.6.50 but also meets a 500 MHz timing target in a 45 nm standard‑cell library. The Fursoy Slice‑1 core, coded entirely in SystemVerilog, implements the full RV64IMAFDV ISA with additional Zba, Zbb, Zicsr and Zifencei extensions, and supports Sv39 virtual memory with demand paging. Its specifications—2‑wide superscalar fetch, a 128‑entry reorder buffer, and a 160‑entry physical register file—place it in the same performance class as many commercial designs, yet the RTL and verification flow remain fully open.
The milestone is more than a technical curiosity. It validates a complete open‑source hardware stack, from RTL synthesis with Yosys to timing analysis with OpenSTA, and demonstrates that community‑driven silicon can achieve frequencies traditionally reserved for proprietary IP. The result could reshape how AI and high‑performance workloads are prototyped, offering a transparent, modifiable alternative to closed‑source solutions.
Fursoy Slice‑1 implements the RV64IMAFDV ISA with Zba, Zbb, Zicsr and Zifencei extensions. The core is a 2‑wide superscalar out‑of‑order engine with a 128‑entry reorder buffer and a 160‑entry physical integer register file, plus 64 vector registers (VLEN = 256, ELEN = 64). Dedicated issue queues include 32‑entry integer and memory queues, a 64‑entry FP register file, and an 8‑entry vector queue. A parallel‑prefix Kogge‑Stone ALU and a three‑stage pipelined multiplier handle arithmetic, while a hardware divider completes division in a single cycle. Sv39 virtual memory is supported through dual‑port DTLB and ITLB structures, each with 64 entries and superpage CAMs. L1I and L1D caches are 16 KiB, 4‑way set‑associative, with non‑blocking features and multiple MSHRs. The SoC integrates a 5‑to‑1 crossbar, 256 MiB DRAM, CLINT, PLIC, UART and a minimal SBI v2.0 firmware.
Physical synthesis on the Nangate 45 nm Open Cell Library reports a total standard‑cell area of 219.43 kGE across eight core blocks, consuming 111,019 cells and 14,836 flip‑flops. Timing closure meets a 2.000 ns clock period, delivering a clean 500 MHz target with zero negative slack in all blocks. The branch unit, ALU, and decode logic all report positive slack, confirming that the design satisfies the frequency goal without aggressive back‑end optimizations. This result demonstrates that a fully open‑source RTL flow—Yosys, sv2v and OpenSTA—can produce a silicon‑ready netlist at a performance level previously reserved for commercial IP.
In Verilator, Fursoy Slice‑1 boots mainline Linux 6.6.50 and BusyBox without kernel panics, oopses, or access faults. Over 89,861,075 simulated cycles, the processor retires 76,236,390 instructions, achieving an IPC of 0.84838. U‑mode commits total 2,227,761, while demand page faults number 2,255 across 1,027 U‑mode threads. The boot log shows successful SBI v2.0 detection, UART initialization at 0x1000_0000, and the launch of /init. No CPI conservation errors are reported, indicating that the cycle‑exact hardware telemetry aligns with the simulated performance model.
The Fursoy achievement lowers the barrier for community‑driven silicon that can handle AI and data‑center workloads. A 500 MHz out‑of‑order core with vector support (VLEN = 256) offers enough compute density for inference kernels while remaining fully transparent in RTL. The open‑source toolchain proves capable of delivering timing‑closed designs, suggesting that future AI accelerators could be built without costly licensing. If the full SoC place‑and‑route and SRAM macro integration succeed, the design could serve as a reference platform for academic research and startups targeting RISC‑V‑based AI processors.
It supports RV64IMAFDV plus Zba, Zbb, Zicsr and Zifencei extensions.
The core meets a 500 MHz target in the Nangate 45 nm standard‑cell library.
No. The simulation reports zero kernel panics, zero oopses and zero access faults during the boot.
With Linux 6.6.50 running flawlessly and timing closure secured at 500 MHz, Fursoy sets a new benchmark for community‑driven processor design. The next hurdle—full SoC place‑and‑route and SRAM integration—will determine whether the core can transition from simulation to silicon. If it does, the open‑source model could accelerate innovation across AI, robotics and quantum‑computing platforms, forcing established vendors to confront a transparent competitor that can deliver comparable performance without licensing fees.
Sources: Hacker News post, GitHub repository https://github.com/FURSOY/fursoy-cpu