Your CPU spends a lot of time waiting for data to arrive from memory. To hide that latency, chips use a trick called prefetching: guess what memory the program will need next, and fetch it early so it's already sitting in the cache when the program asks. Guess right, and things fly. Guess wrong, and you've wasted bandwidth and polluted the cache with junk.

For years now, researchers have been building fancy machine-learning prefetchers — small neural networks that learn access patterns — and reporting that they beat the boring old classical predictors (like "stride" prefetchers that just notice you're walking through memory in fixed-size steps). This paper says: hold on, you're measuring the wrong thing.

Here's the key insight. Any prefetcher has two parts:

• The predictor: guesses which address to fetch.
• The admission gate: decides whether the guess is confident enough to actually issue the fetch.

Classical stride prefetchers in prior comparisons always fire — no gate. The neural ones typically have a confidence threshold and only fire when they're sure. So when the neural version "wins," you can't tell whether it's because the neural model is smarter, or just because it's more selective about when to speak up.

The authors do the experiment the fair way: they bolt the same confidence gate onto both a 257-parameter online MLP (a tiny neural net) and a classical stride predictor, and run them on matched workloads. The result is deflating for the neural-prefetcher crowd:

• On random memory traffic, the neural network is indistinguishable from gated stride.
• On regular streaming patterns, the neural network is actually slower than gated stride.
• The gate itself, however, is genuinely useful — a well-tuned admission policy improves both predictors substantially.

In other words, most of the "AI wins in hardware" story in this space appears to be the admission policy doing the work, not the learned model. The neural network is expensive silicon that duplicates what a much simpler predictor already does — once you stop letting the classical predictor fire indiscriminately.

This matters because hardware real estate is precious. Every transistor spent on an on-chip neural net is transistors not spent on cache, cores, or something else. If the win is really coming from a cheap confidence check, we should build that and skip the neural model.