← Back to News
2026-08-26 RELEASE

MN-Core SDK v0.8 Released

MN-Core SDK v0.8 is released. Here is the release notes.

MN-Core SDK 0.8 : Release Notes

Added

  • Set an upper limit for the number of annealing steps for preset_options at optimization levels O2 and above.
  • Added the io_spec_constraints argument to Context.compile and the input_specs and output_specs properties to CompiledFunction. These APIs enable device-resident values to be passed directly between compiled functions without a host copy or DRAM-to-DRAM layout conversion.
  • Added the whisper_inference advanced example, which demonstrates speech recognition with Whisper and reports its word error rate on LibriSpeech.
  • Added the Context.allocator_state API for inspecting the device memory allocator state.
  • Added the mlsdk.get_all_devices API to list all available devices and their device types.
  • MNCL: Added clCreateCommandQueueWithProperties, including optional Perfetto tracing support with CL_QUEUE_PERFETTO_ENABLE_MNCL for profiling.
  • MNCL: Added program binary queries through CL_PROGRAM_BINARY_SIZES and CL_PROGRAM_BINARIES to obtain MNCL device binary built by clBuildProgram.
  • MNCL: Added host scalar and vector types.
  • MNCL: Added support for multiple kernel entry points that share global values.
  • MNCL: Added new MN-Core 2 device builtins.
  • MNCL: Added MNIR-backend, a backend to use a new intermediate representation for optimization, as an experimental feature.
  • MNCL: Added MNCL Developer’s Manual.

Changed

  • MNCL: IMPORTANT: Programs created with clCreateProgramWithIL, clCreateProgramWithSource or clCreateProgramWithBinary must now be built with clBuildProgram before a kernel is created from them.

Deprecated

  • switch_context feature is deprecated. Although the function still exists, it is a thin dummy now.
  • MNCL: Deprecated clCreateCommandQueue in favor of clCreateCommandQueueWithProperties.

Fixed

  • Fixed the hold and update behavior of the L1BM turnaround register ($lbi) in the MN-Core 2 (GPFN3) emulator, and corrected the corresponding description in the MN-Core 2 Software Developer Manual.
  • Made various performance improvements and bug fixes across MN-Core SDK.

HotFix

When running the whisper_inference example program, apply the following patch first:

cd /opt/pfn/pfcomp/codegen/MLSDK/examples/whisper_inference/
cat << 'EOF' > fix.patch
@@ -16,7 +16,7 @@
     python3 -m venv --system-site-packages ${VENV_DIR}
     source ${VENV_DIR}/bin/activate
     pip install -r ${CURRENT_DIR}/requirements.txt
-    pip install -r ${CURRENT_DIR}/requirements_cpu.txt --index-url https://download.pytorch.org/whl/cpu
+    pip install --force-reinstall -r ${CURRENT_DIR}/requirements_cpu.txt --index-url https://download.pytorch.org/whl/cpu
 else
     source ${VENV_DIR}/bin/activate
 fi
EOF
patch run_whisper_inference.sh < fix.patch