diff --git a/web/packages/new/photos/services/ml/math.ts b/web/packages/new/photos/services/ml/math.ts
index 28b3c1c5eb..7da08b3163 100644
--- a/web/packages/new/photos/services/ml/math.ts
+++ b/web/packages/new/photos/services/ml/math.ts
@@ -11,6 +11,19 @@ export const clamp = (value: number, min: number, max: number) =>
  * the two given vectors.
  *
  * Precondition: The two vectors must be of the same length.
+ *
+ * [Note: Dot product performance]
+ *
+ * In theory, WASM SIMD instructions should give us a huge boost for computing
+ * dot products. In practice, we can get to roughly around the same performance
+ * by using Float32Arrays instead of number[], and letting the JS JIT do the
+ * optimizations for us. (This assertion was made on Chrome on macOS on Sep
+ * 2023, and may not hold in the future).
+ *
+ * We can get an extra 2x speedup over this by using some library that directly
+ * uses the SIMD intrinsics provided by the architecture instead of limiting
+ * itself to the WASM's set. But that requires bundling native code, so as a
+ * tradeoff to avoid complexity we live with leaving that 1x on the table.
  */
 export const dotProduct = (v1: number[], v2: number[]) => {
     if (v1.length != v2.length)
@@ -21,6 +34,16 @@ export const dotProduct = (v1: number[], v2: number[]) => {
     return d;
 };
 
+export const dotProductF32 = (v1: Float32Array, v2: Float32Array) => {
+    if (v1.length != v2.length)
+        throw new Error(`Length mismatch ${v1.length} ${v2.length}`);
+    let d = 0;
+    // eslint-disable-next-line @typescript-eslint/no-non-null-assertion
+    for (let i = 0; i < v1.length; i++) d += v1[i]! * v2[i]!;
+    return d;
+};
+
+
 /**
  * Return the L2-norm ("magnitude") of the given vector.
  *