Inazhar labsbyazhar·Oct 28Choosing the right abstraction for a real AI kernel (CUDA vs Triton vs CUTLASS vs HIP)A common question is: should I learn CUDA or Triton? The better question is: what level of control does this problem actually need? GPU…
InStackademicbyFlagOS·Oct 25FlagOS Cross-Chip GEMM Optimization: Triton Outperforms Native ImplementationsEditor’s Note：GEMM is a core operator in large-model inference, but real workloads often have irregular shapes that are difficult for a…
Eddie Rantsimele·Oct 25Streaming the Weights Pt. 1W4A16 GEMV on Hopper: Triton to the roofline, CUDA for control
Inazhar labsbyazhar·Oct 23How matrix multiplication teaches almost every important GPU optimizationIf you want one workload that teaches coalescing, tiling, registers, shared memory, Tensor Cores, pipelining, occupancy and autotuning, use…
Vijay Goyal·Oct 15Why is my PyTorch model still slow?What I learned optimizing inference with custom Triton kernelsA response icon1A response icon1
Nandana Dileep·Oct 1113 Performance Experiments on One Reranker: The Limits of Faster GPU CodeBenchmarking compilation, custom kernels, batching, and serving
Subrata Goswami·Oct 3LLM Assisted Kernel Development in CuTe DSL, Gluon, Helion and Triton .Humans still have edge in crafting the most performant kernels. This edge is likely to erode within a year or so. Also, in the same spirit…
InGoPenAIbyAadishagrawal·Sep 7Which Thread Holds Which Number?A gentle introduction to Gluon, Triton’s low-level frontend, and the XOR algebra that powers it
