diff --git a/benchmarks/README.md b/benchmarks/README.md
index 72dd5a4a..70c8e234 100644
--- a/benchmarks/README.md
+++ b/benchmarks/README.md
@@ -9,7 +9,7 @@ The `npbench/` suite uses kernels from [npbench](https://github.com/spcl/npbench
 | File | Ufuncs | Dtypes | Sizes/Presets |
 |------|--------|--------|---------------|
 | `micro/bench_micro.py` | 25 unary (`exp`, `log`, `sin`, `cos`, `sqrt`, `cbrt`, etc.) + `arctan2`, `power` | float32, float64 | 10k, 100k, 1M |
-| `npbench/bench_softmax.py` | `exp`, `max`, `sum` | float32 | M (32x8x256x256), L (64x16x448x448) |
+| `npbench/bench_softmax.py` | `exp`, `sum` (`max` runs on stock NumPy) | float32 | M (32x8x256x256), L (64x16x448x448) |
 | `npbench/bench_arc_distance.py` | `sin`, `cos`, `arctan2`, `sqrt` | float64 | M (1M), L (10M) |
 | `npbench/bench_go_fast.py` | `tanh` | float64 | M (6k x 6k), L (20k x 20k) |
 | `npbench/bench_mandelbrot.py` | `abs`, `multiply`, `add` | complex128 | M (250/500), L (833/1000) |
diff --git a/benchmarks/benchmarks/npbench/bench_softmax.py b/benchmarks/benchmarks/npbench/bench_softmax.py
index 41ee7e65..4b7e882c 100644
--- a/benchmarks/benchmarks/npbench/bench_softmax.py
+++ b/benchmarks/benchmarks/npbench/bench_softmax.py
@@ -1,4 +1,7 @@
-"""npbench wrapper: Softmax — mkl_umath ops: exp, max, sum.
+"""npbench wrapper: Softmax — mkl_umath ops: exp, sum.
+
+The kernel also calls np.max, which always runs on stock NumPy: oneMKL has
+no NaN-propagating maximum, so mkl_umath does not provide `maximum`.
 
 Preset sizes from npbench bench_info/softmax.json:
   M: N=32,  H=8,   SM=256   (float32)