diff --git a/core/src/ops/cnn/conv/im2col.rs b/core/src/ops/cnn/conv/im2col.rs
index 59148c8392..c1ca64c5af 100644
--- a/core/src/ops/cnn/conv/im2col.rs
+++ b/core/src/ops/cnn/conv/im2col.rs
@@ -157,8 +157,10 @@ impl EvalOp for Im2Col {
             let r = geometry.out_format.r();
             // Buffer geometry. zero_init for PackedI8K4: the K=4-inner writer skips
             // the K-padding lanes (k..k_aligned), which SMOPA accumulates — they must
-            // be 0. PackedFormat has no K padding; its mn-padding maps to discarded
-            // output rows, so uninitialized is fine (matches prior behaviour).
+            // be 0. PackedFormat has no K padding; its mn-padding lanes are computed
+            // on (then discarded) by the kernel, so the partial last panel still
+            // needs zeroing — garbage there decodes to denormals that stall the fp
+            // pipeline. Done after allocation below.
             let (single_panel_len, buf_align, zero_init) =
                 if let Some(pf) = geometry.out_format.downcast_ref::<PackedFormat>() {
                     (pf.single_panel_len(geometry.k), pf.alignment(), false)
@@ -186,6 +188,8 @@ impl EvalOp for Im2Col {
                     )?;
                     if zero_init {
                         data.as_bytes_mut().fill(0);
+                    } else if n % r != 0 {
+                        data.as_bytes_mut()[(n / r) * panel_bytes..].fill(0);
                     }
                     if n > 0 {
                         dispatch_copy_by_size!(Patcher::patch(dt)(
diff --git a/linalg/src/frame/mmm/scratch.rs b/linalg/src/frame/mmm/scratch.rs
index 4f6ecb8168..bef26d0134 100644
--- a/linalg/src/frame/mmm/scratch.rs
+++ b/linalg/src/frame/mmm/scratch.rs
@@ -381,11 +381,12 @@ impl<TI: LADatum> ScratchSpaceImpl<TI> {
                                         .get_unchecked(..m_remnant),
                                 );
                             }
-                            if cfg!(debug_assertions) {
-                                buf.get_unchecked_mut(m_remnant..)
-                                    .iter_mut()
-                                    .for_each(|x| *x = TI::zero());
-                            }
+                            // The kernel computes on the tail lanes before their
+                            // results are discarded; garbage there decodes to
+                            // denormals and stalls the fp pipeline. Zero them.
+                            buf.get_unchecked_mut(m_remnant..)
+                                .iter_mut()
+                                .for_each(|x| *x = TI::zero());
                             buf.as_ptr()
                         } else {
                             v.as_ptr_unchecked::<TI>().add(down * ker.mr())
@@ -409,11 +410,9 @@ impl<TI: LADatum> ScratchSpaceImpl<TI> {
                                         .get_unchecked(..n_remnant),
                                 );
                             }
-                            if cfg!(debug_assertions) {
-                                buf.get_unchecked_mut(n_remnant..)
-                                    .iter_mut()
-                                    .for_each(|x| *x = TI::zero());
-                            }
+                            buf.get_unchecked_mut(n_remnant..)
+                                .iter_mut()
+                                .for_each(|x| *x = TI::zero());
                             buf.as_ptr()
                         } else {
                             v.as_ptr_unchecked::<TI>().add(right * ker.nr())
@@ -435,11 +434,9 @@ impl<TI: LADatum> ScratchSpaceImpl<TI> {
                                     .get_unchecked(down * ker.mr()..)
                                     .get_unchecked(..m_remnant),
                             );
-                            if cfg!(debug_assertions) {
-                                r.get_unchecked_mut(m_remnant..)
-                                    .iter_mut()
-                                    .for_each(|x| *x = TI::zero());
-                            }
+                            r.get_unchecked_mut(m_remnant..)
+                                .iter_mut()
+                                .for_each(|x| *x = TI::zero());
                             r.as_ptr()
                         } else {
                             rows.as_ptr_unchecked::<TI>().add(down * ker.mr())
@@ -454,11 +451,9 @@ impl<TI: LADatum> ScratchSpaceImpl<TI> {
                                     .get_unchecked(right * ker.nr()..)
                                     .get_unchecked(..n_remnant),
                             );
-                            if cfg!(debug_assertions) {
-                                r.get_unchecked_mut(n_remnant..)
-                                    .iter_mut()
-                                    .for_each(|x| *x = TI::zero());
-                            }
+                            c.get_unchecked_mut(n_remnant..)
+                                .iter_mut()
+                                .for_each(|x| *x = TI::zero());
                             c.as_ptr()
                         } else {
                             cols.as_ptr_unchecked::<TI>().add(right * ker.nr())
@@ -473,9 +468,7 @@ impl<TI: LADatum> ScratchSpaceImpl<TI> {
                         let tile_ptr = store.ptr.offset(tile_offset);
                         let tmp_d_tile =
                             std::slice::from_raw_parts_mut(loc as *mut TI, ker.mr() * ker.nr());
-                        if cfg!(debug_assertions) {
-                            tmp_d_tile.iter_mut().for_each(|t| *t = TI::zero());
-                        }
+                        tmp_d_tile.iter_mut().for_each(|t| *t = TI::zero());
                         for r in 0..m_remnant as isize {
                             for c in 0..n_remnant as isize {
                                 let inner_offset = c * col_byte_stride + r * row_byte_stride;
diff --git a/linalg/src/frame/pack.rs b/linalg/src/frame/pack.rs
index fbdfc053d6..b19dda4c20 100644
--- a/linalg/src/frame/pack.rs
+++ b/linalg/src/frame/pack.rs
@@ -203,6 +203,11 @@ impl PackedFormat {
             );
             if cfg!(debug_assertions) {
                 packed.as_bytes_mut().fill(0u8);
+            } else if mn % self.r != 0 {
+                // The kernel computes on the last panel's padding lanes before
+                // their results are discarded; garbage bytes there decode to
+                // denormals and stall the fp pipeline. Zero the partial panel.
+                packed.as_bytes_mut()[(mn / self.r) * panel_bytes..].fill(0u8);
             }
             dispatch_copy!(Self::pack_t(t.datum_type())(
                 self,
@@ -246,6 +251,11 @@ impl PackedFormat {
             );
             if cfg!(debug_assertions) {
                 packed.as_bytes_mut().fill(0u8);
+            } else if mn % self.r != 0 {
+                // The kernel computes on the last panel's padding lanes before
+                // their results are discarded; garbage bytes there decode to
+                // denormals and stall the fp pipeline. Zero the partial panel.
+                packed.as_bytes_mut()[(mn / self.r) * panel_bytes..].fill(0u8);
             }
             dispatch_copy!(Self::pack_t(t.datum_type())(
                 self,