diff --git a/tools/replay/decode_threads.h b/tools/replay/decode_threads.h
new file mode 100644
index 0000000000..47ed7baa29
--- /dev/null
+++ b/tools/replay/decode_threads.h
@@ -0,0 +1,8 @@
+#pragma once
+
+// Zero means unspecified: preserve FFmpeg's existing default behavior.
+inline int replay_decode_threads(const char *value) {
+  if (value == nullptr) return 0;
+  if (value[0] >= '1' && value[0] <= '4' && value[1] == '\0') return value[0] - '0';
+  return -1;
+}
diff --git a/tools/replay/framereader.cc b/tools/replay/framereader.cc
index a5f0f748c7..2030a9b56a 100644
--- a/tools/replay/framereader.cc
+++ b/tools/replay/framereader.cc
@@ -1,6 +1,8 @@
 #include "tools/replay/framereader.h"
 
 #include <map>
+#include <cstdlib>
+#include "tools/replay/decode_threads.h"
 #include <memory>
 #include <tuple>
 #include <utility>
@@ -152,10 +154,33 @@ bool FFmpegVideoDecoder::open(AVCodecParameters *codecpar, bool hw_decoder) {
     rWarning("No device with hardware decoder found. fallback to CPU decoding.");
   }
 
+  const int slice_threads = replay_decode_threads(std::getenv("ROADSCORE_REPLAY_DECODE_THREADS"));
+  if (slice_threads < 0) {
+    rError("ROADSCORE_REPLAY_DECODE_THREADS must be an integer from 1 to 4");
+    return false;
+  }
+  const bool configure_slices = slice_threads > 0 && hw_device_ctx == nullptr;
+  if (configure_slices) {
+    if (!(decoder->capabilities & AV_CODEC_CAP_SLICE_THREADS)) {
+      rError("Requested software decoder does not support slice threading");
+      return false;
+    }
+    decoder_ctx->thread_count = slice_threads;
+    decoder_ctx->thread_type = FF_THREAD_SLICE;
+  }
+
   if (avcodec_open2(decoder_ctx, decoder, nullptr) < 0) {
     rError("Failed to open codec");
     return false;
   }
+  if (configure_slices) {
+    rInfo("REPLAY_SOFTWARE_DECODER requested_threads=%d threads=%d active_thread_type=%d", slice_threads,
+          decoder_ctx->thread_count, decoder_ctx->active_thread_type);
+    if (decoder_ctx->active_thread_type & FF_THREAD_FRAME) {
+      rError("Frame threading is incompatible with replay's immediate-frame decoder");
+      return false;
+    }
+  }
   return true;
 }
 
diff --git a/tools/replay/tests/test_decode_threads.cc b/tools/replay/tests/test_decode_threads.cc
new file mode 100644
index 0000000000..aee9f281d1
--- /dev/null
+++ b/tools/replay/tests/test_decode_threads.cc
@@ -0,0 +1,12 @@
+#include <cassert>
+#include <initializer_list>
+#include "tools/replay/decode_threads.h"
+int main() {
+ assert(replay_decode_threads(nullptr) == 0);
+ assert(replay_decode_threads("1") == 1);
+ assert(replay_decode_threads("2") == 2);
+ assert(replay_decode_threads("4") == 4);
+ for (const char *value : {"", "0", "5", "-1", "2x", " 2", "2 ", "9999999999999999999"}) {
+   assert(replay_decode_threads(value) == -1);
+ }
+}