diff --git a/pim-os/src/bin/classic_matrix_matrix_multiply.rs b/pim-os/src/bin/classic_matrix_matrix_multiply.rs
new file mode 100644
index 0000000..22aabfd
--- /dev/null
+++ b/pim-os/src/bin/classic_matrix_matrix_multiply.rs
@@ -0,0 +1,23 @@
+#![no_std]
+#![no_main]
+
+extern crate alloc;
+
+use core::fmt::Write;
+use nalgebra::SMatrix;
+use pim_os::{pim::vector::F16x1, uart::Uart0};
+
+#[no_mangle]
+pub extern "C" fn main() {
+    let matrices0 = [SMatrix::<F16x1, 8, 8>::zeros(); 512];
+    let matrices1 = [SMatrix::<F16x1, 8, 8>::zeros(); 512];
+
+    for _ in 0..100 {
+        for i in 0..512 {
+            let matrix2 = matrices0[i] * matrices1[i];
+            core::hint::black_box(matrix2);
+        }
+    }
+
+    writeln!(Uart0, "Done").unwrap();
+}
diff --git a/pim-os/src/m5op.rs b/pim-os/src/m5op.rs
index 734c2c1..b9a8a1c 100644
--- a/pim-os/src/m5op.rs
+++ b/pim-os/src/m5op.rs
@@ -1,4 +1,3 @@
-use aarch64_cpu::asm::barrier;
 use core::arch::global_asm;
 
 global_asm!(include_str!("m5op.s"));
@@ -25,61 +24,3 @@ pub fn dump_stats() {
 pub fn dump_reset_stats() {
     unsafe { m5_dump_reset_stats() }
 }
-
-// const M5OPS_ADDR: u64 = 0x10010000;
-
-// const EXIT_ADDR: *const u64 = (M5OPS_ADDR + (0x21 << 8)) as _;
-// const DUMP_STATS_ADDR: *const u64 = (M5OPS_ADDR + (0x40 << 8)) as _;
-// const RESET_STATS_ADDR: *const u64 = (M5OPS_ADDR + (0x41 << 8)) as _;
-// const DUMP_RESET_STATS_ADDR: *const u64 = (M5OPS_ADDR + (0x42 << 8)) as _;
-// const WORK_BEGIN_ADDR: *const u64 = (M5OPS_ADDR + (0x5A << 8)) as _;
-// const WORK_END_ADDR: *const u64 = (M5OPS_ADDR + (0x5B << 8)) as _;
-// // const CHECKPOINT_ADDR: *const u64 = (M5OPS_ADDR + (0x43 << 8)) as _;
-
-// pub fn exit() {
-//     unsafe {
-//         core::ptr::read_volatile(EXIT_ADDR);
-//     }
-
-//     barrier::dsb(barrier::SY);
-// }
-
-// pub fn dump_stats() {
-//     unsafe {
-//         core::ptr::read_volatile(DUMP_STATS_ADDR);
-//     }
-
-//     barrier::dsb(barrier::SY);
-// }
-
-// pub fn reset_stats() {
-//     unsafe {
-//         core::ptr::read_volatile(RESET_STATS_ADDR);
-//     }
-
-//     barrier::dsb(barrier::SY);
-// }
-
-// pub fn dump_reset_stats() {
-//     unsafe {
-//         core::ptr::read_volatile(DUMP_RESET_STATS_ADDR);
-//     }
-
-//     barrier::dsb(barrier::SY);
-// }
-
-// pub fn work_begin() {
-//     unsafe {
-//         core::ptr::read_volatile(WORK_BEGIN_ADDR);
-//     }
-
-//     barrier::dsb(barrier::SY);
-// }
-
-// pub fn work_end() {
-//     unsafe {
-//         core::ptr::read_volatile(WORK_END_ADDR);
-//     }
-
-//     barrier::dsb(barrier::SY);
-// }
diff --git a/pim-os/src/pim/config.rs b/pim-os/src/pim/config.rs
index c82a3cd..9de82c8 100644
--- a/pim-os/src/pim/config.rs
+++ b/pim-os/src/pim/config.rs
@@ -1,5 +1,8 @@
 use aarch64_cpu::asm::barrier;
-use core::{arch::asm, ptr::write_volatile};
+use core::{
+    arch::asm,
+    ptr::{addr_of_mut, write_volatile},
+};
 
 #[link_section = ".pim_config"]
 static mut PIM_CONFIG_REGION: [u8; 0x4000] = [0; 0x4000];
@@ -8,11 +11,17 @@ pub fn write(s: &str) {
     unsafe {
         let mut index = 0;
         for &byte in s.as_bytes() {
-            write_volatile((&mut PIM_CONFIG_REGION as *mut u8).offset(index), byte as _);
+            write_volatile(
+                (addr_of_mut!(PIM_CONFIG_REGION) as *mut u8).offset(index),
+                byte as _,
+            );
             barrier::dsb(barrier::SY);
             index += 1;
         }
-        write_volatile((&mut PIM_CONFIG_REGION as *mut u8).offset(index), b'\0');
+        write_volatile(
+            (addr_of_mut!(PIM_CONFIG_REGION) as *mut u8).offset(index),
+            b'\0',
+        );
 
         // PIM_CONFIG_REGION[..s.len()].copy_from_slice(s.as_bytes());
         // PIM_CONFIG_REGION[s.len()] = b'\0';