Data movement
Modern processors can execute arithmetic faster than they can feed it; pointer-rich state repeatedly crosses cache and memory boundaries.
Menu
01 / THE THESIS
The next bottleneck is moving and coordinating state.
Compute keeps getting faster. Feeding it—and coordinating irregular work across the machine—is becoming the harder systems problem.
Next: Product thesisDivergence
Dense compute hardware excels on regular kernels, but graphs, branches, and sparse control leave execution resources idle or waiting.
Fragmented execution
One application crosses CPU, GPU, and specialized runtimes—each with different kernels, schedulers, and tuning rules.
THE RESPONSE
NLM makes computation local.
Instead of moving instructions and data to a distant processor, NLM lets computation emerge where the data already lives.
GRAPH-NATIVE EXECUTION
- 01
Programs are represented as nodes and links.
- 02
An active pair triggers a local rewrite.
- 03
Independent active pairs execute in parallel.
- 04
Numerical primitives use the same triggering model.
