From the archive
Thursday, 12 November
Folded from 20 sources that day.
Sorted by quality and by how many sources covered it, the same for everyone.
-
Fast by Default: Compile-Time Heuristics for Helion Kernels
Helion now picks kernel configs at compile time from IR and dataflow, skipping autotuning. It reports 1.18x over hand-written FLA Triton and 1.27x over vLLM's CUDA operators on H100.
-
Training a 1T-Parameter MoE on EU public infrastructure with PyTorch and TorchTitan
The EUROPA consortium pretrained a 951.9B-parameter MoE on the JUPITER supercomputer with TorchTitan. Throughput rose from 90 to about 1,300 tokens/s per GPU on 384 GH200 GPUs.