diff --git a/.gitignore b/.gitignore
index 6a58583c..5ebd9b58 100644
--- a/.gitignore
+++ b/.gitignore
@@ -16,3 +16,6 @@ Cargo.lock
 
 # some scene files are not checked into version control
 local_scenes
+
+# profiling
+profile.json
diff --git a/Cargo.toml b/Cargo.toml
index 578be356..e6a98435 100644
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -6,3 +6,7 @@ members = ["clovers", "clovers-cli"]
 [profile.release]
 codegen-units = 1
 lto = "fat"
+
+[profile.profiling]
+inherits = "release"
+debug = true
diff --git a/Justfile b/Justfile
index 0fa31ec7..43b44fc2 100644
--- a/Justfile
+++ b/Justfile
@@ -21,3 +21,8 @@ all-scenes:
   for scene in $(ls scenes/ |grep json); \
   do just cli -s 1 --input scenes/$scene --output renders/$DATE/${scene%.json}.png; \
   done;
+
+# Profiling helper
+profile *ARGS:
+  cargo build --profile profiling; \
+  samply record -- ./target/profiling/clovers-cli {{ARGS}}
diff --git a/clovers-cli/Cargo.toml b/clovers-cli/Cargo.toml
index c8240161..04546764 100644
--- a/clovers-cli/Cargo.toml
+++ b/clovers-cli/Cargo.toml
@@ -18,12 +18,12 @@ clovers = { path = "../clovers", features = [
 ], default-features = false }
 
 # External
-clap = { version = "4.4.1", features = ["std", "derive"] }
+clap = { version = "4.4.7", features = ["std", "derive"] }
 human_format = "1.0.3"
 humantime = "2.1.0"
 image = { version = "0.24.7", features = ["png"], default-features = false }
 img-parts = "0.3.0"
-indicatif = { version = "0.17.6", features = [
+indicatif = { version = "0.17.7", features = [
   "rayon",
 ], default-features = false }
 palette = { version = "0.7.3", features = ["serializing"] }
@@ -31,9 +31,9 @@ rand = { version = "0.8.5", features = [
   "small_rng",
   "getrandom",
 ], default-features = false }
-rayon = "1.7.0"
-serde = { version = "1.0.188", features = ["derive"], default-features = false }
+rayon = "1.8.0"
+serde = { version = "1.0.190", features = ["derive"], default-features = false }
 serde_json = { version = "1.0", features = ["alloc"], default-features = false }
-time = { version = "0.3.28", default-features = false }
-tracing = "0.1.37"
+time = { version = "0.3.30", default-features = false }
+tracing = "0.1.40"
 tracing-subscriber = { version = "0.3.17", features = ["time"] }
diff --git a/clovers-cli/src/draw_cpu.rs b/clovers-cli/src/draw_cpu.rs
index 389ab6d1..7b5af8cb 100644
--- a/clovers-cli/src/draw_cpu.rs
+++ b/clovers-cli/src/draw_cpu.rs
@@ -9,72 +9,80 @@ use rand::rngs::SmallRng;
 use rand::{Rng, SeedableRng};
 use rayon::prelude::*;
 use scenes::Scene;
-use std::time::Duration;
 
 /// The main drawing function, returns a `Vec<Srgb>` as a pixelbuffer.
 pub fn draw(opts: RenderOpts, scene: &Scene) -> Vec<Srgb<u8>> {
-    // Progress bar
-    let pixels = (opts.width * opts.height) as u64;
-    let bar = ProgressBar::new(pixels);
+    let width = opts.width as usize;
+    let height = opts.height as usize;
+    let bar = progress_bar(&opts);
 
-    if opts.quiet {
-        bar.set_draw_target(ProgressDrawTarget::hidden())
-    } else {
-        bar.set_style(ProgressStyle::default_bar().template(
-            "Elapsed: {elapsed_precise}\nPixels:  {bar} {pos}/{len}\nETA:     {eta_precise}",
-        ).unwrap());
-        bar.enable_steady_tick(Duration::from_millis(100));
-    }
-
-    let black: Srgb<u8> = Srgb::new(0, 0, 0);
-    let mut pixelbuffer = vec![black; pixels as usize];
-
-    pixelbuffer
-        .par_iter_mut()
-        .enumerate()
-        .for_each(|(index, pixel)| {
-            // Enumerate gives us an usize, opts.width and opts.height are u32
-            // Most internal functions expect a Float, perform conversions
-            let x = (index % (opts.width as usize)) as Float;
-            let y = (index / (opts.width as usize)) as Float;
-            let width = opts.width as Float;
-            let height = opts.height as Float;
-
-            // Initialize a thread-local random number generator
+    let pixelbuffer: Vec<Srgb<u8>> = (0..height)
+        .into_par_iter()
+        .map(|row_index| {
             let mut rng = SmallRng::from_entropy();
-
-            // Initialize a mutable base color for the pixel
-            let mut color: LinSrgb = LinSrgb::new(0.0, 0.0, 0.0);
-
-            if opts.normalmap {
-                // If we are rendering just a normalmap, make it quick and early return
-                let u = x / width;
-                let v = y / height;
-                let ray: Ray = scene.camera.get_ray(u, v, &mut rng);
-                color = normal_map(&ray, scene, &mut rng);
-                let color: Srgb = color.into_color();
-                *pixel = color.into_format();
-                return;
-            }
-            // Otherwise, do a regular render
-
-            // Multisampling for antialiasing
-            for _sample in 0..opts.samples {
-                if let Some(s) = sample(scene, x, y, width, height, &mut rng, opts.max_depth) {
-                    color += s
+            let mut row = Vec::with_capacity(width);
+            for index in 0..width {
+                let index = index + row_index * width;
+                if opts.normalmap {
+                    row.push(render_pixel_normalmap(scene, &opts, index, &mut rng));
+                } else {
+                    row.push(render_pixel(scene, &opts, index, &mut rng));
                 }
             }
-            color /= opts.samples as Float;
-            // Gamma / component transfer function
-            let color: Srgb = color.into_color();
-            *pixel = color.into_format();
-
             bar.inc(1);
-        });
+            row
+        })
+        .flatten()
+        .collect();
 
     pixelbuffer
 }
 
+// Render a single pixel, including possible multisampling
+fn render_pixel(scene: &Scene, opts: &RenderOpts, index: usize, rng: &mut SmallRng) -> Srgb<u8> {
+    let (x, y, width, height) = index_to_params(opts, index);
+    let mut color: LinSrgb = LinSrgb::new(0.0, 0.0, 0.0);
+    for _sample in 0..opts.samples {
+        if let Some(s) = sample(scene, x, y, width, height, rng, opts.max_depth) {
+            color += s
+        }
+    }
+    color /= opts.samples as Float;
+    let color: Srgb = color.into_color();
+    let color: Srgb<u8> = color.into_format();
+    color
+}
+
+// Render a single pixel in normalmap mode
+fn render_pixel_normalmap(
+    scene: &Scene,
+    opts: &RenderOpts,
+    index: usize,
+    rng: &mut SmallRng,
+) -> Srgb<u8> {
+    let (x, y, width, height) = index_to_params(opts, index);
+    let color: LinSrgb = sample_normalmap(scene, x, y, width, height, rng);
+    let color: Srgb = color.into_color();
+    let color: Srgb<u8> = color.into_format();
+    color
+}
+
+// Get a single sample for a single pixel in the scene, normalmap mode.
+fn sample_normalmap(
+    scene: &Scene,
+    x: Float,
+    y: Float,
+    width: Float,
+    height: Float,
+    rng: &mut SmallRng,
+) -> LinSrgb {
+    let u = x / width;
+    let v = y / height;
+    let ray: Ray = scene.camera.get_ray(u, v, rng);
+    let color = normal_map(&ray, scene, rng);
+    color.into_color()
+}
+
 /// Get a single sample for a single pixel in the scene. Has slight jitter for antialiasing when multisampling.
 fn sample(
     scene: &Scene,
@@ -88,11 +96,31 @@ fn sample(
     let u = (x + rng.gen::<Float>()) / width;
     let v = (y + rng.gen::<Float>()) / height;
     let ray: Ray = scene.camera.get_ray(u, v, rng);
-    let new_color: Xyz<E> = colorize(&ray, scene, 0, max_depth, rng);
-    let new_color: Xyz = new_color.adapt_into();
-    let new_color: LinSrgb = new_color.into_color_unclamped();
-    if new_color.red.is_finite() && new_color.green.is_finite() && new_color.blue.is_finite() {
-        return Some(new_color);
+    let color: Xyz<E> = colorize(&ray, scene, 0, max_depth, rng);
+    let color: Xyz = color.adapt_into();
+    let color: LinSrgb = color.into_color_unclamped();
+    if color.red.is_finite() && color.green.is_finite() && color.blue.is_finite() {
+        return Some(color);
     }
     None
 }
+
+fn index_to_params(opts: &RenderOpts, index: usize) -> (Float, Float, Float, Float) {
+    let x = (index % (opts.width as usize)) as Float;
+    let y = (index / (opts.width as usize)) as Float;
+    let width = opts.width as Float;
+    let height = opts.height as Float;
+    (x, y, width, height)
+}
+
+fn progress_bar(opts: &RenderOpts) -> ProgressBar {
+    let bar = ProgressBar::new(opts.height as u64);
+    if opts.quiet {
+        bar.set_draw_target(ProgressDrawTarget::hidden())
+    } else {
+        bar.set_style(ProgressStyle::default_bar().template(
+            "Elapsed:   {elapsed_precise}\nRows:      {bar} {pos}/{len}\nRemaining: {eta_precise}",
+        ).unwrap());
+    }
+    bar
+}
diff --git a/clovers/Cargo.toml b/clovers/Cargo.toml
index 86ae4580..08caecbe 100644
--- a/clovers/Cargo.toml
+++ b/clovers/Cargo.toml
@@ -20,15 +20,15 @@ traces = ["tracing"]
 [dependencies]
 enum_dispatch = "0.3.12"
 gltf = { version = "1.3.0", optional = true }
-nalgebra = { version = "0.32.3", features = ["libm"], default-features = false }
+nalgebra = { version = "0.32.3" }
 palette = { version = "0.7.3", features = ["serializing"] }
 rand = { version = "0.8.5", features = ["small_rng"], default-features = false }
 rand_distr = "0.4.3"
-serde = { version = "1.0.188", features = [
+serde = { version = "1.0.190", features = [
   "derive",
 ], default-features = false, optional = true }
 stl_io = { version = "0.7.0", optional = true }
-tracing = { version = "0.1.37", optional = true }
+tracing = { version = "0.1.40", optional = true }
 
 [dev-dependencies]
 criterion = "0.4.0"