diff --git a/cpp/CMakeLists.txt b/cpp/CMakeLists.txt
index 32852705a1..a4a4aed084 100644
--- a/cpp/CMakeLists.txt
+++ b/cpp/CMakeLists.txt
@@ -1980,7 +1980,9 @@ elseif(CMAKE_CXX_COMPILER_ID STREQUAL "GNU" OR CMAKE_CXX_COMPILER_ID STREQUAL "C
     # For ARM architecture, as a hack, ensure that char is signed
     message(STATUS "ARM architecture detected: adding -fsigned-char flag")
     set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsigned-char")
-  else()
+  elseif(${CMAKE_SYSTEM_PROCESSOR} MATCHES "(x86_64|amd64|AMD64|i386|i486|i586|i686|x86)")
+    # -mfpmath=sse only exists on x86, so non-x86 ISAs (loongarch64, riscv64, ppc64le, mips64,
+    # ...) must not get it - they hard-error with "unrecognized command-line option".
     # KataGo wants IEEE-compliant float math.
     # On x86-64 SSE-based IEEE float math is already the ABI default (x87 excess precision cannot occur),
     # but mandate it explicitly anyway.
diff --git a/cpp/tests/testsearchcommon.cpp b/cpp/tests/testsearchcommon.cpp
index 6cd6936427..87d8bdc823 100644
--- a/cpp/tests/testsearchcommon.cpp
+++ b/cpp/tests/testsearchcommon.cpp
@@ -203,6 +203,24 @@ NNEvaluator* TestSearchCommon::startNNEval(
   const string& modelName = modelFile;
   const string homeDataDirOverride = "";
   ConfigParser cfg;
+#if defined(USE_EIGEN_BACKEND)
+  //The Eigen backend only implements the NHWC float32 path and hard-errors on anything else
+  //(see eigenbackend.cpp), but the tests and the runsearchtests*.sh scripts are written for the
+  //GPU backends and pass NCHW and/or FP16. Normalize here, at the one choke point every NN test
+  //goes through, rather than fixing up each call site.
+  if(!inputsUseNHWC) {
+    std::cout << "Backend is Eigen, ignoring args and forcing inputsUseNHWC=true" << std::endl;
+    inputsUseNHWC = true;
+  }
+  if(!useNHWC) {
+    std::cout << "Backend is Eigen, ignoring args and forcing useNHWC=true" << std::endl;
+    useNHWC = true;
+  }
+  if(useFP16) {
+    std::cout << "Backend is Eigen, ignoring args and forcing useFP16=false" << std::endl;
+    useFP16 = false;
+  }
+#endif
   //NHWC layout is no longer a generic NNEvaluator option; only the CUDA backend reads it (off cfg).
   //Route the test's useNHWC param into a cudaUseNHWC override so it still drives the CUDA layout.
   cfg.overrideKey("cudaUseNHWC", useNHWC ? "true" : "false");
diff --git a/cpp/tests/testtrainingwrite.cpp b/cpp/tests/testtrainingwrite.cpp
index 80d12d87e8..3a979f8548 100644
--- a/cpp/tests/testtrainingwrite.cpp
+++ b/cpp/tests/testtrainingwrite.cpp
@@ -24,6 +24,19 @@ static NNEvaluator* startNNEval(
   bool debugSkipNeuralNet = modelFile == "/dev/null";
   const string homeDataDirOverride = "";
   ConfigParser cfg;
+#if defined(USE_EIGEN_BACKEND)
+  //The Eigen backend only implements the NHWC float32 path and hard-errors on NCHW
+  //(see eigenbackend.cpp). These tests default to useNHWC=false, which the GPU backends accept
+  //but Eigen does not, so normalize here at the one choke point they all go through.
+  if(!inputsUseNHWC) {
+    cout << "Backend is Eigen, ignoring args and forcing inputsUseNHWC=true" << endl;
+    inputsUseNHWC = true;
+  }
+  if(!useNHWC) {
+    cout << "Backend is Eigen, ignoring args and forcing useNHWC=true" << endl;
+    useNHWC = true;
+  }
+#endif
   //NHWC layout is no longer a generic NNEvaluator option; only the CUDA backend reads it (off cfg).
   //Route the test's useNHWC param into a cudaUseNHWC override so it still drives the CUDA layout.
   cfg.overrideKey("cudaUseNHWC", useNHWC ? "true" : "false");