AI & ML Practical Magic

Our computers are way slower than they should be because they're hardwired to think time only goes one way.

arXiv · March 17, 2026 · 2603.13750

Paul Borrill

AI-generated illustration

The Takeaway

By removing the assumption that a 'cause' must always happen before an 'effect' in network chip communication, researchers eliminated a massive 350% performance delay. This suggests that the way we think about time in technology is actually causing a digital traffic jam.

From the abstract

The datacenter industry is converging on SmartNIC-based resource management. Wave (Humphries et al., ASPLOS '25) demonstrates the practical feasibility of offloading kernel thread scheduling, memory management, and RPC stacks to the ARM cores of Intel's Mount Evans Infrastructure Processing Unit (IPU). The engineering is careful and the results are honest: without Wave's PCIe latency mitigations, offloaded workloads degrade by 350%.We argue that this 350% degradation is not an engineering proble