diff --git a/src/felix86/common/global.cpp b/src/felix86/common/global.cpp
index 95db4c293..7a2ab32c3 100644
--- a/src/felix86/common/global.cpp
+++ b/src/felix86/common/global.cpp
@@ -98,7 +98,7 @@ void ProcessGlobals::initialize() {
 
     perf = std::make_unique<Perf>();
 
-    cas128_lock = 0;
+    memset(cas128_locks, 0, sizeof(cas128_locks));
 
     // HACK: Don't clear as they get shared per mount namespace
     // TODO: proper mount namespacing when we need it
diff --git a/src/felix86/common/global.hpp b/src/felix86/common/global.hpp
index d4d0ec2ea..377a90a11 100644
--- a/src/felix86/common/global.hpp
+++ b/src/felix86/common/global.hpp
@@ -49,7 +49,7 @@ struct ProcessGlobals {
     std::unique_ptr<Perf> perf;
 
     // For cmpxchg16b
-    u32 cas128_lock = 0;
+    u32 cas128_locks[256];
 
     // TODO: this isn't per CLONE_VM but per mount namespace
     // But we don't care for now
diff --git a/src/felix86/v2/handlers.cpp b/src/felix86/v2/handlers.cpp
index 164b2137d..f6f635321 100644
--- a/src/felix86/v2/handlers.cpp
+++ b/src/felix86/v2/handlers.cpp
@@ -10684,10 +10684,21 @@ FAST_HANDLE(CMPXCHG16B) {
         biscuit::Label spinloop, writeloop;
         biscuit::GPR lock_address = rec.scratch();
         biscuit::GPR lock = rec.scratch();
-        as.LI(lock_address, (u64)&g_process_globals.cas128_lock);
+        as.LI(lock_address, (u64)&g_process_globals.cas128_locks);
+        // We will pick one of 256 different spinlocks based on a hash created by our address
+        // This means that if two cmpxchg16b target the same address they will spin on the same lock
+        // but if they target a different address they will likely get a different lock, which should decrease
+        // lock contention
+        constexpr u32 knuth_hash = 2654435761u;
+        as.LI(mem1, knuth_hash);
+        as.SRLI(mem0, address, 4); // shift out low bits since they are 0 to get a better hash
+        as.MULW(mem0, mem0, mem1);
+        as.ANDI(mem0, mem0, 0xFF);
+        static_assert(sizeof(g_process_globals.cas128_locks) == 256 * sizeof(u32));
+        as.SH2ADD(lock_address, mem0, lock_address);
 
-        as.Bind(&spinloop);
         as.LI(lock, 1);
+        as.Bind(&spinloop);
         as.AMOSWAP_W(Ordering::AQRL, lock, lock, lock_address);
         as.BNEZ(lock, &spinloop);