Minerva is a CPU core that currently implements the RISC-V RV32IM instruction set. Its microarchitecture is described in plain Python code using the nMigen toolbox.

Minerva currently requires Python 3.6+ and nMigen on its master branch.

python setup.py install
python build.py > minerva.v

To use Minerva, you need to wire the following ports to minerva_cpu:

• clk
• rst
• ibus__*
• dbus__*
• external_interrupt

The microarchitecture of Minerva is largely inspired by the LatticeMico32 processor.

Minerva is pipelined on 6 stages:

1. Address The address of the next instruction is calculated and sent to the instruction cache.
2. Fetch The instruction is read from memory.
3. Decode The instruction is decoded, and operands are either fetched from the register file or bypassed from the pipeline. Branches are predicted by the static branch predictor.
4. Execute Simple instructions such as arithmetic and logical operations are completed at this stage.
5. Memory More complicated instructions such as loads, stores and shifts require a second execution stage.
6. Writeback Results produced by the instructions are written back to the register file.

The two L1 caches are write-through, meaning that writes are done to both the cache and the underlying memory hierarchy.

Miss penalties are reduced by restarting execution as soon as the missing word is available, without waiting for the complete line refill. Furthermore, cache refills always request the missing word first. By combining these two policies, Minerva is able to resume its execution after a cache miss in only two clock cycles in the best case, regardless of line size.

The L1 data cache is coupled to a write buffer. Store transactions are in this case done to the write buffer instead of the data bus. This enables stores to proceed in one clock cycle if the buffer isn't full, without having to wait for the bus transaction to complete. Store transactions are then completed in the background as the write buffer gets emptied to the data bus.

Minerva is able to operate at 100MHz on a Xilinx XC7A35-2 FPGA, with both caches enabled.

The following parameters can be used to configure the Minerva core.

In no particular order:

• RV64I
• Floating Point Unit
• Stateful branch prediction
• ...

If you are interested in sponsoring new features or improvements, get in touch at contact [at] lambdaconcept.com .

Minerva is released under the permissive two-clause BSD license. See LICENSE file for full copyright and license information.