diff --git a/pipeline/tests/powdr_std.rs b/pipeline/tests/powdr_std.rs
index 0254b7bf1..55e3d7771 100644
--- a/pipeline/tests/powdr_std.rs
+++ b/pipeline/tests/powdr_std.rs
@@ -58,6 +58,14 @@ fn split_gl_test() {
     gen_estark_proof(make_simple_prepared_pipeline(f));
 }
 
+#[cfg(feature = "plonky3")]
+#[test]
+#[ignore = "Too slow"]
+fn split_bb_test() {
+    let f = "std/split_bb_test.asm";
+    test_plonky3_with_backend_variant::<BabyBearField>(f, vec![], BackendVariant::Composite);
+}
+
 #[test]
 #[ignore = "Too slow"]
 fn arith_test() {
diff --git a/std/machines/split/mod.asm b/std/machines/split/mod.asm
index 866a5d488..bfd89c2c7 100644
--- a/std/machines/split/mod.asm
+++ b/std/machines/split/mod.asm
@@ -1,5 +1,6 @@
 mod split_bn254;
 mod split_gl;
+mod split_bb;
 
 use std::utils::cross_product;
 
@@ -21,4 +22,4 @@ machine ByteCompare with
 
     col fixed latch = [1]*;
     col fixed operation_id = [0]*;
-}
\ No newline at end of file
+}
diff --git a/std/machines/split/split_bb.asm b/std/machines/split/split_bb.asm
new file mode 100644
index 000000000..0f7285434
--- /dev/null
+++ b/std/machines/split/split_bb.asm
@@ -0,0 +1,77 @@
+use std::prelude::Query;
+use super::ByteCompare;
+
+// Splits an arbitrary field element into two u16s, on the BabyBear field.
+machine SplitBB(byte_compare: ByteCompare) with
+    latch: RESET,
+    // Allow this machine to be connected via a permutation
+    call_selectors: sel,
+{
+    operation split in_acc -> output_low, output_high;
+
+    // Latch and operation ID
+    col fixed RESET(i) { if i % 4 == 3 { 1 } else { 0 } };
+
+    // 1. Decompose the input into bytes
+
+    // The byte decomposition of the input, in little-endian order
+    // and shifted forward by one (to use the last row of the
+    // previous block)
+    // A hint is provided because automatic witness generation does not
+    // understand step 3 to figure out that the byte decomposition is unique.
+    let select_byte: fe, int -> fe = |input, byte| std::convert::fe((std::convert::int(input) >> (byte * 8)) & 0xff);
+    col witness bytes;
+    query |i| {
+        std::prover::provide_value(bytes, i, select_byte(std::prover::eval(in_acc'), (i + 1) % 4));
+    };
+    // Puts the bytes together to form the input
+    col witness in_acc;
+    // Factors to multiply the bytes by
+    col fixed FACTOR(i) { 1 << (((i + 1) % 4) * 8) };
+
+    in_acc' = (1 - RESET) * in_acc + bytes * FACTOR;
+
+    // 2. Build the output, packing chunks of 2 bytes (i.e., 16 bit) into a field element
+    col witness output_low, output_high;
+    col fixed FACTOR_OUTPUT_LOW = [0x100, 0, 0, 1]*;
+    col fixed FACTOR_OUTPUT_HIGH = [0, 1, 0x100, 0]*;
+    output_low' = (1 - RESET) * output_low + bytes * FACTOR_OUTPUT_LOW;
+    output_high' = (1 - RESET) * output_high + bytes * FACTOR_OUTPUT_HIGH;
+
+    // 3. Check that the byte decomposition does not overflow
+    //
+    //    Skipping this step would work but it wouldn't be sound, because
+    //    the 4-byte decomposition could overflow, since the BabyBear
+    //    prime 2**31 - 2**27 + 1 is smaller than 2^32.
+    //
+    //    The approach is to compare the byte decomposition with that of
+    //    the maximum possible value (0x78000000) byte by byte,
+    //    from most significant to least significant (i.e., going backwards).
+    //    A byte can only be larger than that of the max value if any previous
+    //    byte has been smaller.
+
+    // This is an example for input 0x77ffffff:
+    // Row     RESET   bytes   BYTES_MAX  lt      was_lt  gt
+    // -1      0x1     0xff    0x0        0x0     0x1     0x1
+    //  0      0x0     0xff    0x0        0x0     0x1     0x1
+    //  1      0x0     0xff    0x0        0x0     0x1     0x1
+    //  2      0x0     0x77    0x78       0x1     0x1     0x0  # 0x77 < 0x78, so now greater bytes are allowed
+
+    // Bytes of the maximum value, in little endian order, rotated by one
+    col fixed BYTES_MAX = [0, 0, 0x78, 0]*;
+
+    // Compare the current byte with the corresponding byte of the maximum value.
+    col witness lt;
+    col witness gt;
+    link => (lt, gt) = byte_compare.run(bytes, BYTES_MAX);
+
+    // Compute whether the current or any previous byte has been less than
+    // the corresponding byte of the maximum value.
+    // This moves *backward* from the second to last row.
+    col witness was_lt;
+    was_lt = RESET' * lt + (1 - RESET') * (was_lt' + lt - was_lt' * lt);
+
+    // If any byte is larger, but no previous byte was smaller, the byte
+    // decomposition has overflowed and should be rejected.
+    gt * (1 - was_lt) = 0;
+}
diff --git a/test_data/std/split_bb_test.asm b/test_data/std/split_bb_test.asm
new file mode 100644
index 000000000..fd9e0b272
--- /dev/null
+++ b/test_data/std/split_bb_test.asm
@@ -0,0 +1,49 @@
+use std::machines::split::ByteCompare;
+use std::machines::split::split_bb::SplitBB;
+
+machine Main with degree: 65536 {
+    reg pc[@pc];
+    reg X0[<=];
+    reg X1[<=];
+    reg X2[<=];
+    reg low;
+    reg high;
+
+    ByteCompare byte_compare;
+    SplitBB split_machine(byte_compare);
+
+    instr split X0 -> X1, X2 link ~> (X1, X2) = split_machine.split(X0);
+
+    instr assert_eq X0, X1 {
+        X0 = X1
+    }
+
+    function main {
+
+        // Min value
+        // Note that this has two byte decompositions, 0 and p = 0x78000001.
+        // The second would lead to a different split value, but should be ruled
+        // out by the overflow check.
+        low, high <== split(0);
+        assert_eq low, 0;
+        assert_eq high, 0;
+
+        // Max value
+        // On BabyBear, this is 0x78000000.
+        low, high <== split(-1);
+        assert_eq low, 0;
+        assert_eq high, 0x7800;
+
+        // Max low value
+        low, high <== split(0x77ffffff);
+        assert_eq low, 0xffff;
+        assert_eq high, 0x77ff;
+
+        // Some other value
+        low, high <== split(0x42abcdef);
+        assert_eq low, 0xcdef;
+        assert_eq high, 0x42ab;
+
+        return;
+    }
+}