ISA efficiency comparison

[mratsim]

This adds an instruction count cycle to SP1 to ease comparison between Valida ISA and RISC-V ISA.

This is interesting due to the following

zkVMs	ISA	Proof System
Jolt	RISC-V	Spartan (sumcheck, multilinear)
SP1	RISC-V	Plonky3 (FRI, univariate)
Valida	Valida ISA (RISC-like)	Plonky3 (FRI, univariate)

Video

keccak-ISA-efficiency-bench.mp4

Instructions count:

Valida (native): 15289
SP1 (native): 16800 keccak, 24100 total
SP1 (precompile): 706 keccak, 7916 total

What can we conclude from this?

In general we cannot compare 2 zkvms on raw instruction count and infer efficiency, this can be seen with SP1 keccak precompile with 3x less instructions yet slower proof generation for a single keccak than native execution.

However, for RISC-like instructions (so no complex super instructions like Keccak), and on the same proof system (Plonky3), less instructions usually mean less work for the prover.