diff --git a/Cargo.lock b/Cargo.lock
index 1e7bcf06..33939daa 100644
--- a/Cargo.lock
+++ b/Cargo.lock
@@ -1426,6 +1426,7 @@ dependencies = [
  "num",
  "ordered-float",
  "predicates",
+ "rand",
  "rayon",
  "regex",
  "rust-htslib",
diff --git a/Cargo.toml b/Cargo.toml
index f9d1f655..5bd3d940 100644
--- a/Cargo.toml
+++ b/Cargo.toml
@@ -52,6 +52,7 @@ gzp = "0.11.3"
 niffler = {version = "2.5.0", default-features = false, features = ["gz"]}
 burn = { version = "0.12", optional = true, features = ["candle"] } # "wgpu",
 num = "0.4.3"
+rand = "0.8.5"
 #polars = "0.38"
 
 [build-dependencies]
diff --git a/src/cli/qc_opts.rs b/src/cli/qc_opts.rs
index f21127e3..1b68369a 100644
--- a/src/cli/qc_opts.rs
+++ b/src/cli/qc_opts.rs
@@ -21,6 +21,9 @@ pub struct QcOpts {
     /// maximum number of reads to use in the ACF calculation
     #[clap(long, default_value = "10000")]
     pub acf_max_reads: usize,
+    /// After sampling the first "acf-max-reads" randomly sample one of every "acf-sample-rate" reads and replace one of the previous reads at random.
+    #[clap(long, default_value = "100")]
+    pub acf_sample_rate: f32,
     /// In the output include a measure of the number of m6A events per MSPs of a given size.
     /// The output format is: "m6a_per_msp_size\t{m6A count},{MSP size},{is a FIRE}\t{count}"
     /// e.g. "m6a_per_msp_size\t35,100,false\t100"
diff --git a/src/subcommands/qc.rs b/src/subcommands/qc.rs
index 688b7dc4..450d558d 100644
--- a/src/subcommands/qc.rs
+++ b/src/subcommands/qc.rs
@@ -4,7 +4,9 @@ use crate::utils::bio_io;
 use anyhow::Result;
 use itertools::Itertools;
 use ordered_float::OrderedFloat;
+use rand::prelude::*;
 use std::collections::HashMap;
+use std::collections::VecDeque;
 use std::io::Write;
 
 // set the precision of the floats to be saved and printed
@@ -56,13 +58,11 @@ pub struct QcStats<'a> {
     pub rq: HashMap<OrderedFloat<f32>, i64>,
     // m6a per msp size: (msp size, m6a count, is a FIRE element), number of times seen
     pub m6a_per_msp_size: HashMap<M6aPerMsp, i64>,
-    // add the m6a acf
-    pub m6a_acf: HashMap<i64, f64>,
-    // pub
-    pub acf_read_count: usize,
     // m6a starts for acf
-    m6a_acf_starts: Vec<f64>,
-    //
+    m6a_acf_starts: VecDeque<Vec<f64>>,
+    // times m6as have been sampled at random for acf
+    sampled: usize,
+    // the qc options for printing
     qc_opts: &'a QcOpts,
     // phasing information
     phased_reads: HashMap<String, i64>,
@@ -85,9 +85,8 @@ impl<'a> QcStats<'a> {
             cpg_count: HashMap::new(),
             m6a_per_msp_size: HashMap::new(),
             rq: HashMap::new(),
-            m6a_acf: HashMap::new(),
-            acf_read_count: 0,
-            m6a_acf_starts: Vec::new(),
+            m6a_acf_starts: VecDeque::new(),
+            sampled: 0,
             qc_opts,
             phased_reads: HashMap::new(),
             phased_bp: HashMap::new(),
@@ -96,13 +95,7 @@ impl<'a> QcStats<'a> {
 
     pub fn add_read_to_stats(&mut self, fiber: &fiber::FiberseqData) {
         // add auto-correlation of m6a
-        if self.qc_opts.acf
-            && self.acf_read_count < self.qc_opts.acf_max_reads
-            && fiber.m6a.starts.len() >= self.qc_opts.acf_min_m6a
-        {
-            self.add_m6a_starts_for_acf(fiber);
-            self.acf_read_count += 1;
-        }
+        self.add_m6a_starts_for_acf(fiber);
 
         self.full_read_stats(fiber);
         self.add_basemod_stats(fiber);
@@ -114,11 +107,47 @@ impl<'a> QcStats<'a> {
 
     /// converts the m6A calls into a boolean vector for the ACF calculation
     fn add_m6a_starts_for_acf(&mut self, fiber: &fiber::FiberseqData) {
+        // skip conditions
+        if !self.qc_opts.acf || fiber.m6a.starts.len() < self.qc_opts.acf_min_m6a {
+            return;
+        }
+
+        // test if we should skip or not based on length and random sampling
+        let rand_float: f32 = random();
+        let sample = rand_float < 1.0 / self.qc_opts.acf_sample_rate;
+        if !(self.m6a_acf_starts.len() <= self.qc_opts.acf_max_reads || sample) {
+            return;
+        };
+
+        // note how many times we have sampled
+        if sample {
+            self.sampled += 1;
+        }
+
+        // add the m6a to the working queue
         let mut m6a_vec: Vec<f64> = vec![0.0; fiber.record.seq_len()];
         for m6a in fiber.m6a.starts.iter().flatten() {
             m6a_vec[*m6a as usize] = 1.0;
         }
-        self.m6a_acf_starts.extend(m6a_vec);
+
+        // if we have sampled enough that all reads are random replace
+        // a random previous read with the current read
+        if sample && self.sampled > self.qc_opts.acf_max_reads {
+            let idx = thread_rng().gen_range(0..self.m6a_acf_starts.len());
+            self.m6a_acf_starts[idx] = m6a_vec;
+            log::trace!(
+                "Replaced read at index {} after the {}th sample",
+                idx,
+                self.sampled
+            );
+            return;
+        }
+
+        // otherwise add to the end while constraining the size of the queue
+        self.m6a_acf_starts.push_back(m6a_vec);
+        if self.m6a_acf_starts.len() > self.qc_opts.acf_max_reads {
+            self.m6a_acf_starts.pop_front();
+        }
     }
 
     fn add_ranges(&mut self, fiber: &fiber::FiberseqData) {
@@ -233,10 +262,14 @@ impl<'a> QcStats<'a> {
         if !self.qc_opts.acf {
             return Ok(());
         }
-
         log::info!("Calculating m6A auto-correlation.");
         let acf = crate::utils::acf::acf_par(
-            &self.m6a_acf_starts,
+            &self
+                .m6a_acf_starts
+                .iter()
+                .flatten()
+                .map(|f| *f)
+                .collect::<Vec<f64>>(),
             Some(self.qc_opts.acf_max_lag),
             false,
         )?;