Ascend 950DT is the fifth accelerator family Hoid targets, and the first where we had to write a collective ourselves.
What the backend required
CANN’s pipeline scheduling model is not a drop-in for the stream model the other backends share, so the runtime grew a scheduling adapter rather than a set of conditionals in the hot path. Wide expert parallelism needed a hand-rolled collective: the vendor primitive assumed a topology our deployment shape doesn’t have.
The Pareto curve
We ran GLM-5 across MI355X, B200 and 950DT at matched accuracy, sweeping batch size and expert parallelism. The interesting result isn’t a single peak number — it’s that the crossover point between the three moves with sequence length, and it moves a long way.