Grey PVM Benchmark Results¶

Grey's RISC-V recompiler vs PolkaVM's compiler backend, measured on 8 workloads covering compute, memory, crypto, and host call patterns.

Methodology¶

Gas metering model: Pipeline gas metering (POLKAVM_DEFAULT_COST_MODEL=full-l1-hit). This charges gas per basic block based on instruction count, matching the JAR v0.8.0 specification. Both Grey and PolkaVM use synchronous gas metering under this model.

Why compile+execute matters: In JAM, every work-package arrives as a PVM blob that must be compiled and executed from scratch. There is no persistent code cache across invocations -- each validator independently compiles the blob before execution. The compile+execute measurement reflects this real-world cost. We report exec-only numbers separately for engineering insight, but the full pipeline is the metric that determines block processing time.

Benchmark fairness: - Both VMs compile from the same source (Rust, compiled to RISC-V, transpiled to PVM for Grey / compiled to PolkaVM format for PolkaVM) - Grey recompiler uses mprotect-based bounds checking (always enabled on x86-64 Linux) - PolkaVM uses generic sandbox with POLKAVM_ALLOW_EXPERIMENTAL=1 - Criterion statistical framework, 100 samples per benchmark (10 for ecrecover) - Same host machine, same gas limit, same workload parameters

Reproduction:

POLKAVM_ALLOW_EXPERIMENTAL=1 POLKAVM_DEFAULT_COST_MODEL=full-l1-hit \
  cargo bench -p grey-bench

Results: Compile + Execute (full pipeline)¶

This is the number that matters for blockchain performance.

Grey wins all 8 benchmarks. The advantage ranges from 1.01x (sieve) to 3.97x (hostcall), with a geometric mean speedup of ~1.9x.

The primary driver is compilation speed: Grey's single-pass x86-64 code generator compiles PVM bytecode approximately 3x faster than PolkaVM's compiler backend. For short programs (blake2b, keccak) where compilation dominates total time, this translates directly into large end-to-end wins. For long-running programs (fib, sieve) where execution dominates, the two VMs converge.

Results: Execution Only¶

Compilation performed in setup (untimed). Measures native code quality in isolation.

Grey wins 7 of 8 benchmarks on exec-only. PolkaVM's only win is ecrecover (1.09x). Grey generates notably better native code for hash functions (blake2b 2.0x, keccak 2.0x) and host call dispatch (3.9x). The hostcall result is notable: PolkaVM's compiled code (3,213 us) is actually slower than its own interpreter (2,526 us) on this workload, suggesting high overhead in its compiler's host call dispatch path.

Results: Interpreters¶

Grey's interpreter wins 6 of 8. PolkaVM wins ed25519 (1.6x) and ecrecover (1.1x).

Workload Descriptions¶

Environment¶

• Date: 2026-03-30
• Grey: javm recompiler (mprotect bounds checking)
• PolkaVM: v0.32.0, generic sandbox, full-l1-hit cost model
• Rust: 1.94.1, release mode
• Hardware: Intel Core i9-13900K