Fursoy Slice-1 RISC-V Core Release
Overview
Fursoy Slice-1 is a clean-slate 64-bit out-of-order (OoO) superscalar RISC-V processor implemented in SystemVerilog. The core implements the RV64IMAFDV instruction set extension, including Zba, Zbb, Zicsr, and Zifencei. It supports Sv39 virtual memory, privilege separation (Machine/Supervisor/User), and boots mainline Linux 6.6.50 with BusyBox userspace. The design is optimized for a 45nm Nangate standard-cell library, achieving 500 MHz timing closure across its primary functional blocks.
Microarchitecture
The core utilizes a 2-wide superscalar pipeline for fetch, decode, rename, and in-order retirement. Key structural features include:
- Reorder Buffer (ROB): 128-entry depth with 4-bit wrap-safe epoch tags.
- Register File: 160-entry Physical Integer Register File (PRF) with 128 speculative free-list entries.
- Branch Prediction: 8 branch checkpoints for single-cycle speculative Register Alias Table (RAT) rollback.
- Issue Queues: Dedicated queues for Integer (32-entry), Memory (32-entry), Floating Point (64-entry), and Vector operations (8-entry).
- Execution Units: Parallel-prefix 64-bit Kogge-Stone ALU, a 3-stage pipelined multiplier, and a dedicated hardware divider.
Memory and SoC Integration
Fursoy Slice-1 features a non-blocking memory subsystem designed to handle complex page table walks and cache coherency. The L1D cache is 16 KiB, 4-way set-associative, and 8-banked, supporting 16-entry Load/Store queues.
Notably, the hardware Page Table Walker (PTW) implements non-destructive L1D dirty snoop coherency. This allows page-table writes residing in the L1D to be observed directly without requiring an eviction or a full cache flush. The SoC integration includes a 5-to-1 priority crossbar, CLINT, PLIC, and a 16550 UART, managed by a custom SBI v2.0 M-mode firmware.
Physical Synthesis and Verification
The design was evaluated using Yosys 0.33, sv2v v0.0.13, and OpenSTA 3.1.0. Across 8 core functional blocks, the design occupies approximately 219 kGE and meets a 2.000 ns clock period (500 MHz) in the Nangate 45nm library.
Verification was performed using a Verilator-based simulation environment. The core successfully completed over 89 million cycles, retiring over 76 million instructions with an IPC of approximately 0.85. The project includes a cycle-exact CPI conservation breakdown, which tracks the attribution of every cycle to specific pipeline states, such as branch recovery, memory stalls, or productive graduation.
Community Discussion
Community feedback highlighted the technical difficulty of implementing an out-of-order processor as a solo project. The author noted that the most significant debugging challenge involved a livelock during Linux boot, where a CSR read-modify-write instruction on the mip register incorrectly fed back into the RMW logic, causing the core to latch an interrupt signal and trap the kernel in a loop.
Regarding the development process, the author noted:
"For folks who have built Sv39 cores or open-source ASIC flows: do you prefer letting the PTW snoop dirty L1D lines directly like this, or forcing a clean/drain path? Also curious what people would tackle next on a single-core tree like this—an inclusive L2 cache or jumping straight to dual-core coherency."
While some observers expressed a desire to see the design implemented on physical silicon rather than just simulated in Verilog, the project provides a complete, open-source reference for a modern, Linux-capable RISC-V core.