DeepSeek Releases DeepGEMM Ascend With Runtime Kernel Compilation and Fused Mega MoE
Summary
DeepSeek releases DeepGEMM Ascend, using DeepJIT runtime kernel compilation to avoid CUDA builds at installation and a fused Mega MoE kernel that overlaps NVLink transfers with tensor-core computation.
Key Points
- DeepSeek releases DeepGEMM Ascend on September 30, 2026, alongside locality-domain optimizations.
- DeepGEMM compiles kernels at runtime through DeepJIT, eliminating CUDA compilation during installation.
- Mega MoE fuses expert-parallel dispatch, two linear layers, SwiGLU and combine operations into one kernel that overlaps NVLink communication with tensor-core computation.