diff --git a/crates/prover/benches/merkle.rs b/crates/prover/benches/merkle.rs
index c6735cdb6..0ff92235e 100644
--- a/crates/prover/benches/merkle.rs
+++ b/crates/prover/benches/merkle.rs
@@ -1,59 +1,42 @@
 #![feature(iter_array_chunks)]
 
-use criterion::Criterion;
+use criterion::{criterion_group, criterion_main, Criterion, Throughput};
+use itertools::Itertools;
+use num_traits::Zero;
+use stwo_prover::core::backend::simd::SimdBackend;
+use stwo_prover::core::backend::{CPUBackend, Col};
+use stwo_prover::core::fields::m31::BaseField;
+use stwo_prover::core::vcs::blake2_merkle::Blake2sMerkleHasher;
+use stwo_prover::core::vcs::ops::MerkleOps;
 
-#[cfg(target_arch = "x86_64")]
-pub fn cpu_merkle(c: &mut criterion::Criterion) {
-    use itertools::Itertools;
-    use num_traits::Zero;
-    use stwo_prover::core::backend::avx512::AVX512Backend;
-    use stwo_prover::core::backend::{CPUBackend, Col};
-    use stwo_prover::core::fields::m31::BaseField;
-    use stwo_prover::core::vcs::ops::MerkleOps;
-    use stwo_prover::platform;
+const LOG_N_ROWS: u32 = 16;
 
-    const N_COLS: usize = 1 << 8;
-    const LOG_SIZE: u32 = 16;
-    let cols = (0..N_COLS)
-        .map(|_| {
-            (0..(1 << LOG_SIZE))
-                .map(|_| BaseField::zero())
-                .collect::<Vec<_>>()
-        })
-        .collect::<Vec<_>>();
+const LOG_N_COLS: u32 = 8;
 
+fn bench_blake2s_merkle<B: MerkleOps<Blake2sMerkleHasher>>(c: &mut Criterion, id: &str) {
+    let col: Col<B, BaseField> = (0..1 << LOG_N_ROWS).map(|_| BaseField::zero()).collect();
+    let cols = (0..1 << LOG_N_COLS).map(|_| col.clone()).collect_vec();
+    let col_refs = cols.iter().collect_vec();
     let mut group = c.benchmark_group("merkle throughput");
-    group.throughput(criterion::Throughput::Elements((N_COLS << LOG_SIZE) as u64));
-    group.throughput(criterion::Throughput::Bytes(
-        (N_COLS << (LOG_SIZE + 2)) as u64,
-    ));
-    group.bench_function("cpu merkle", |b| {
-        b.iter(|| {
-            CPUBackend::commit_on_layer(LOG_SIZE, None, &cols.iter().collect_vec());
-        })
+    group.throughput(Throughput::Elements(1 << (LOG_N_COLS + LOG_N_ROWS)));
+    group.throughput(Throughput::Bytes(4 << (LOG_N_COLS + LOG_N_ROWS)));
+    group.bench_function(&format!("{id} merkle"), |b| {
+        b.iter_with_large_drop(|| B::commit_on_layer(LOG_N_ROWS, None, &col_refs))
     });
+}
 
-    if !platform::avx512_detected() {
-        return;
+fn blake2s_merkle_benches(c: &mut Criterion) {
+    #[cfg(target_arch = "x86_64")]
+    if stwo_prover::platform::avx512_detected() {
+        use stwo_prover::core::backend::avx512::AVX512Backend;
+        bench_blake2s_merkle::<AVX512Backend>(c, "avx");
     }
-    let cols = (0..N_COLS)
-        .map(|_| {
-            (0..(1 << LOG_SIZE))
-                .map(|_| BaseField::zero())
-                .collect::<Col<AVX512Backend, BaseField>>()
-        })
-        .collect::<Vec<_>>();
-
-    group.bench_function("avx merkle", |b| {
-        b.iter(|| {
-            AVX512Backend::commit_on_layer(LOG_SIZE, None, &cols.iter().collect_vec());
-        })
-    });
+    bench_blake2s_merkle::<SimdBackend>(c, "simd");
+    bench_blake2s_merkle::<CPUBackend>(c, "cpu");
 }
 
-#[cfg(target_arch = "x86_64")]
-criterion::criterion_group!(
-    name=merkle;
+criterion_group!(
+    name = benches;
     config = Criterion::default().sample_size(10);
-    targets=cpu_merkle);
-criterion::criterion_main!(merkle);
+    targets = blake2s_merkle_benches);
+criterion_main!(benches);
diff --git a/crates/prover/src/core/backend/simd/blake2s.rs b/crates/prover/src/core/backend/simd/blake2s.rs
index 77b9a1b3f..58b74b7ae 100644
--- a/crates/prover/src/core/backend/simd/blake2s.rs
+++ b/crates/prover/src/core/backend/simd/blake2s.rs
@@ -110,7 +110,8 @@ fn rotate<const N: u32>(x: u32x16) -> u32x16 {
     (x >> N) | (x << (u32::BITS - N))
 }
 
-#[inline]
+// `inline(always)` can cause code parsing errors for wasm: "locals exceed maximum".
+#[cfg_attr(not(target_arch = "wasm32"), inline(always))]
 fn round(v: &mut [u32x16; 16], m: [u32x16; 16], r: usize) {
     v[0] += m[SIGMA[r][0] as usize];
     v[1] += m[SIGMA[r][2] as usize];