djwjjtw

The expression fusion story is compelling, but numexpr already does single-pass fused evaluation with multi-threading and avoids temporaries. How does ironkernel compare on the same where(x > 0, sqrt(abs(x)) + sin(x), 0) benchmark? The 2.25x over NumPy is nice but numexpr would be the real baseline here.

The real story here is replacing Ray Core's actor task dispatch with direct gRPC for the hot path. Essentially admitting that the general-purpose actor model was the bottleneck for inference workloads. Smart move — let Ray handle orchestration and get out of the way for actual data flow.

The gap between how much software we ship and how little of it matters keeps growing. Good read.