diff --git a/Cargo.lock b/Cargo.lock
index 33db1df..dffd5e9 100644
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -867,6 +867,10 @@ dependencies = [
  "serde_json",
 ]
 
+[[package]]
+name = "wasmi_benches_mandelbrot"
+version = "0.0.0"
+
 [[package]]
 name = "wasmi_benches_matrix_mul"
 version = "0.0.0"
@@ -894,6 +898,10 @@ version = "0.0.0"
 name = "wasmi_benches_sort"
 version = "0.0.0"
 
+[[package]]
+name = "wasmi_benches_spectralnorm"
+version = "0.0.0"
+
 [[package]]
 name = "wasmi_benches_tiny_keccak"
 version = "0.0.0"
diff --git a/Cargo.toml b/Cargo.toml
index b5e58c4..aeac44c 100644
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -20,6 +20,8 @@ members = [
     "cases/word_count",
     "cases/compression",
     "cases/json_parse",
+    "cases/spectralnorm",
+    "cases/mandelbrot",
 ]
 resolver = "2"
 
diff --git a/cases/mandelbrot/Cargo.toml b/cases/mandelbrot/Cargo.toml
new file mode 100644
index 0000000..d4b1fa0
--- /dev/null
+++ b/cases/mandelbrot/Cargo.toml
@@ -0,0 +1,14 @@
+[package]
+name = "wasmi_benches_mandelbrot"
+version.workspace = true
+publish.workspace = true
+authors.workspace = true
+repository.workspace = true
+rust-version.workspace = true
+edition.workspace = true
+
+[lib]
+path = "lib.rs"
+crate-type = ["cdylib"]
+
+[dependencies]
diff --git a/cases/mandelbrot/lib.rs b/cases/mandelbrot/lib.rs
new file mode 100644
index 0000000..05c1b0c
--- /dev/null
+++ b/cases/mandelbrot/lib.rs
@@ -0,0 +1,79 @@
+extern crate alloc;
+
+use alloc::{boxed::Box, vec};
+
+/// Maximum number of escape-time iterations evaluated per pixel.
+///
+/// Pixels inside the Mandelbrot set never escape and therefore run the full
+/// budget, which is what makes this benchmark compute bound.
+const MAX_ITER: u32 = 1000;
+
+/// Left edge of the rendered region on the real axis.
+const MIN_X: f64 = -2.0;
+/// Bottom edge of the rendered region on the imaginary axis.
+const MIN_Y: f64 = -1.25;
+/// Side length of the (square) rendered region in the complex plane.
+///
+/// The region spans `[-2.0, 0.5] × [-1.25, 1.25]`, a 2.5 × 2.5 window centered
+/// on `(-0.75, 0.0)` that captures the whole classic Mandelbrot shape. Keeping
+/// it square means each pixel maps to a square cell for any `size`.
+const SPAN: f64 = 2.5;
+
+#[repr(C)]
+pub struct MandelbrotData {
+    /// Per-pixel escape iteration counts, laid out row-major as `size × size`.
+    ///
+    /// There is no random input to seed: the image is fully determined by the
+    /// fixed view region and [`MAX_ITER`], so the buffer is simply (re)computed
+    /// from the pixel coordinates on every [`run`].
+    iterations: Box<[u32]>,
+    /// Width and height of the (quadratic) render area in pixels.
+    size: usize,
+}
+
+#[unsafe(no_mangle)]
+pub extern "C" fn setup(size: usize) -> Box<MandelbrotData> {
+    let pixels = size.checked_mul(size).expect("render area overflows usize");
+    Box::new(MandelbrotData {
+        iterations: vec![0; pixels].into_boxed_slice(),
+        size,
+    })
+}
+
+#[unsafe(no_mangle)]
+pub extern "C" fn teardown(_: Box<MandelbrotData>) {}
+
+#[unsafe(no_mangle)]
+pub extern "C" fn run(data: &mut MandelbrotData) {
+    let size = data.size;
+    // Distance between adjacent pixels in the complex plane.
+    let step = SPAN / size as f64;
+    for py in 0..size {
+        let cy = MIN_Y + py as f64 * step;
+        let row = &mut data.iterations[py * size..(py + 1) * size];
+        for (px, cell) in row.iter_mut().enumerate() {
+            let cx = MIN_X + px as f64 * step;
+            // Escape-time iteration of z := z² + c starting from z = 0.
+            let mut zx = 0.0;
+            let mut zy = 0.0;
+            let mut iter = 0;
+            while iter < MAX_ITER {
+                let zx2 = zx * zx;
+                let zy2 = zy * zy;
+                if zx2 + zy2 > 4.0 {
+                    break;
+                }
+                zy = 2.0 * zx * zy + cy;
+                zx = zx2 - zy2 + cx;
+                iter += 1;
+            }
+            *cell = iter;
+        }
+    }
+}
+
+#[unsafe(no_mangle)]
+pub extern "C" fn output(data: &MandelbrotData) -> u64 {
+    // Sum every escape count so the optimizer cannot elide the computation.
+    data.iterations.iter().map(|&it| it as u64).sum()
+}
diff --git a/cases/mandelbrot/out.wasm b/cases/mandelbrot/out.wasm
new file mode 100644
index 0000000..325678a
Binary files /dev/null and b/cases/mandelbrot/out.wasm differ
diff --git a/cases/spectralnorm/Cargo.toml b/cases/spectralnorm/Cargo.toml
new file mode 100644
index 0000000..d8646cc
--- /dev/null
+++ b/cases/spectralnorm/Cargo.toml
@@ -0,0 +1,12 @@
+[package]
+name = "wasmi_benches_spectralnorm"
+version.workspace = true
+publish.workspace = true
+authors.workspace = true
+repository.workspace = true
+rust-version.workspace = true
+edition.workspace = true
+
+[lib]
+path = "lib.rs"
+crate-type = ["cdylib"]
diff --git a/cases/spectralnorm/lib.rs b/cases/spectralnorm/lib.rs
new file mode 100644
index 0000000..44ebd9e
--- /dev/null
+++ b/cases/spectralnorm/lib.rs
@@ -0,0 +1,140 @@
+extern crate alloc;
+
+use alloc::{boxed::Box, vec, vec::Vec};
+
+/// Number of power-iteration steps performed per `run`.
+///
+/// This is the classic constant used by the "spectral norm" benchmark and is
+/// enough for the dominant eigenvalue estimate to converge.
+const ITERATIONS: usize = 10;
+
+#[repr(C)]
+pub struct SpectralNormData {
+    /// Working vector (left operand of the power iteration).
+    u: Box<[f64]>,
+    /// Working vector (right operand of the power iteration).
+    v: Box<[f64]>,
+    /// Scratch space holding `A * x` between the two half-steps.
+    tmp: Box<[f64]>,
+    /// Randomized starting vector, copied into `u` at the start of each `run`
+    /// so that every run performs identical, deterministic work.
+    u_init: Box<[f64]>,
+    /// Estimated spectral norm from the last run.
+    result: f64,
+}
+
+/// Deterministic LCG, identical in spirit to the one used by the `nbody` case.
+struct Lcg {
+    state: u64,
+}
+
+impl Lcg {
+    fn new(seed: u64) -> Self {
+        Self { state: seed }
+    }
+
+    fn next_u64(&mut self) -> u64 {
+        self.state = self.state.wrapping_mul(6364136223846793005).wrapping_add(1);
+        self.state
+    }
+
+    fn next_f64(&mut self) -> f64 {
+        const SCALE: f64 = 1.0 / ((1u64 << 53) as f64);
+        ((self.next_u64() >> 11) as f64) * SCALE
+    }
+
+    fn range(&mut self, min: f64, max: f64) -> f64 {
+        min + (max - min) * self.next_f64()
+    }
+}
+
+/// Element `A(i, j)` of the infinite symmetric matrix used by the benchmark.
+///
+/// The denominator is computed with integer arithmetic (as in the canonical
+/// "spectral norm" benchmark) before a single conversion to `f64`. We use `u64`
+/// rather than `usize` on purpose: on `wasm32` `usize` is 32-bit, so the
+/// `ij * (ij + 1)` product would silently overflow for larger dimensions.
+#[inline]
+fn matrix(i: usize, j: usize) -> f64 {
+    let i = i as u64;
+    let j = j as u64;
+    let ij = i + j;
+    1.0 / ((ij * (ij + 1) / 2 + i + 1) as f64)
+}
+
+/// Computes `out = A * src`.
+fn mul_a(src: &[f64], out: &mut [f64]) {
+    let n = src.len();
+    for (i, out_i) in out.iter_mut().enumerate() {
+        let mut sum = 0.0;
+        for (j, &s) in src.iter().enumerate().take(n) {
+            sum += matrix(i, j) * s;
+        }
+        *out_i = sum;
+    }
+}
+
+/// Computes `out = Aᵀ * src` (using `A(j, i)` instead of `A(i, j)`).
+fn mul_at(src: &[f64], out: &mut [f64]) {
+    let n = src.len();
+    for (i, out_i) in out.iter_mut().enumerate() {
+        let mut sum = 0.0;
+        for (j, &s) in src.iter().enumerate().take(n) {
+            sum += matrix(j, i) * s;
+        }
+        *out_i = sum;
+    }
+}
+
+#[unsafe(no_mangle)]
+pub extern "C" fn setup(len: usize) -> Box<SpectralNormData> {
+    let mut rng = Lcg::new(len as u64);
+    // Randomized, strictly positive starting vector. Any vector that is not
+    // orthogonal to the dominant eigenvector converges, so the exact values do
+    // not matter; positivity simply guarantees a well-behaved start.
+    let u_init: Vec<f64> = (0..len).map(|_| rng.range(0.5, 1.5)).collect();
+    Box::new(SpectralNormData {
+        u: vec![0.0; len].into_boxed_slice(),
+        v: vec![0.0; len].into_boxed_slice(),
+        tmp: vec![0.0; len].into_boxed_slice(),
+        u_init: u_init.into_boxed_slice(),
+        result: 0.0,
+    })
+}
+
+#[unsafe(no_mangle)]
+pub extern "C" fn teardown(_: Box<SpectralNormData>) {}
+
+#[unsafe(no_mangle)]
+pub extern "C" fn output(data: &SpectralNormData) -> f64 {
+    data.result
+}
+
+#[unsafe(no_mangle)]
+pub extern "C" fn run(data: &mut SpectralNormData) {
+    // Reset the working vector to the randomized start so each run is identical.
+    data.u.copy_from_slice(&data.u_init);
+
+    let u: &mut [f64] = &mut data.u;
+    let v: &mut [f64] = &mut data.v;
+    let tmp: &mut [f64] = &mut data.tmp;
+    // Power iteration: repeatedly apply `AᵀA` to refine the dominant eigenvector.
+    for _ in 0..ITERATIONS {
+        // v = AᵀA * u
+        mul_a(u, tmp);
+        mul_at(tmp, v);
+        // u = AᵀA * v
+        mul_a(v, tmp);
+        mul_at(tmp, u);
+    }
+
+    // After the loop `u = AᵀA·v`, so the Rayleigh quotient
+    // ||A||₂ ≈ sqrt(uᵀv / vᵀv) estimates the largest singular value.
+    let mut v_bv = 0.0;
+    let mut vv = 0.0;
+    for (&ui, &vi) in u.iter().zip(v.iter()) {
+        v_bv += ui * vi;
+        vv += vi * vi;
+    }
+    data.result = (v_bv / vv).sqrt();
+}
diff --git a/cases/spectralnorm/out.wasm b/cases/spectralnorm/out.wasm
new file mode 100644
index 0000000..e614814
Binary files /dev/null and b/cases/spectralnorm/out.wasm differ