ESP32C6's Wasmi performance test report

[vhexl]

This is ESP32C6's WASMI performance test report reference for somebody.

wasmi v1.1 on ESP32-C6: Quick Benchmark Summary

Target: ESP32-C6, RISC-V RV32IMAC @ 160 MHz, no FPU
Runtime: wasmi v1.1 (interpretation-only, no WASM proposals)
Compiler: Rust nightly, opt-level=s, LTO=fat, build-std
wasmi config: default-features = false, features = ["prefer-btree-collections"]

Core Benchmarks

Op	ns/op	Notes
i32.add	489	Hardware ALU
i32.mul	495	Hardware multiplier
i32.div_u	558	Hardware divider
f32.add	846	Soft-float (no FPU), 1.73× vs i32
f32.mul	821	1.68× vs i32
f32.div	1,123	2.30× vs i32
if/else branch	1,098	Per-iteration
wasm-internal call	2,892	2-level chain + i32.add payload
i32.store8	1,221 KB/s	Linear memory write + bounds check
i32.load8_u	993 KB/s	Linear memory read + bounds check

Overhead

Metric	Value
Engine init (Engine + Module + Store + instantiate)	3,216 µs
Empty () → () cross-boundary call	5,689 ns
Host function call (Taylor sin)	17,381 ns
└ Boundary crossing alone (controlled experiment)	3,420 ns
└ Actual computation (Taylor sin, native code)	4,580 ns

Volume

Resource	Used	Capacity
Flash	~668 KB	8 MB (8.2%)
RAM	~442 KB	512 KB (86%)
└ wasmi heap	80 KB	—
└ Stack (ESP-IDF default)	353 KB	Can reduce to 64 KB

Key Takeaways

1. Batching host calls wins big: Replacing 300 per-iteration host calls with a single batch call saves 33% (3,052 → 2,032 µs).
2. Boundary crossing is ~3.4 µs per call — comparable to the host-side computation itself on a 160 MHz MCU.
3. Fixed-point recommended: Without FPU, f32 ops are 1.7–2.3× slower than i32. Use u32/i32 fixed-point in wasm hot loops.
4. Stack is the low-hanging fruit: Reducing ESP-IDF stack from 353→64 KB frees ~290 KB RAM.
5. LTO is critical: lto = 'fat' + codegen-units = 1 is essential to fit wasmi + HAL on 8 MB Flash.
6. Reproducible: Two independent flash cycles — 32/33 metrics identical, single f32.mul 1-unit drift (0.12%).

Build Environment

# rust-toolchain.toml
channel = "nightly"
components = ["rust-src", "llvm-tools-preview"]
targets = ["riscv32imac-unknown-none-elf"]

# Cargo.toml — wasmi
wasmi = { version = "1.1", default-features = false,
          features = ["prefer-btree-collections"] }

# Cargo.toml — runtime
esp-hal = "~1.1.0"
esp-rtos = "0.3.0"        # Embassy async runtime
esp-alloc = "0.10.0"      # 80 KB heap
embassy-time = "0.5.1"    # µs timer

Full Report

wasmi-esp32c6-benchmark.md — raw data, flash/RAM maps, controlled experiment details, reproducibility data.

---

wasmi-esp32c6-benchmark.md
wasmi_perf.zip

[Robbepop]

Hi @vhexl and thanks for your report!

Since this is for embedded Wasmi users I wonder how much better is Wasmi v2.0.0-beta.7 since it has a ton of new useful features for embedded users in particular:

• A new validate crate feature allows to turn off Wasm validation shrinking Wasmi by up to 350 kB.
  • Users have to pre-validate the input Wasm themselves and use Wasmi's Module::new_unchecked API.
• A new memory64 crate feature can toggle Wasm memory64 at compile time, giving decent perf and minimal size wins on embedded targets when turned off.
• Wasmi now uses slightly different build configs when opt-level = "s" is used to shrink even further.
• A new auto-dispatch feature can be used to always use the best possible dispatching method on each target.
  • Since the ESPc6 is using a RISC-V chip, Wasmi v2.0.0-beta.7 should be able to use the new highly efficient tail-call dispatch for it with auto-dispatch enabled.
• Tons of other optimizations and a whole architecture change should boost performance further.