We Deleted Our GPU Speed Claims Instead of Fixing Them

Some users have reported the same thing: a healthy, modern graphics card, drivers up to date, everything else on the machine using it happily — and our app saying it cannot find a compatible GPU.

This post is not an announcement that we fixed it. It is about what we did in the meantime, because that decision is the more interesting one.

A working GPU and a usable GPU are different things

It is reasonable to assume that if your card runs games and your driver tools report no problems, any application should be able to use it. That assumption does not hold for machine-learning work.

To run a model on a GPU, an application needs a runtime path to it: a component that translates the model's operations into work the card understands. On Windows that is usually DirectML; on NVIDIA hardware it can be CUDA; on Apple Silicon it is CoreML. The application has to ship that path, the path has to load at startup, and it has to be compatible with the specific driver and hardware in front of it.

Break any link in that chain and the result is identical from the outside: the framework reports that no usable device is available. Your card is fine. The route to it is not.

We do ship the DirectML path on Windows, and the app tries it at startup and falls back to CPU when it does not validate. On some machines it does not validate, and we do not yet have a confident account of why. That is an honest description of where we are.

The part we could fix immediately

Here is what was true while all of that was unresolved: our app and our website carried copy about GPU acceleration. One line promised "the fastest possible speed." Another explained slowness in terms of the user's hardware.

Read those next to the reports above and the problem is obvious. For a user whose GPU never attaches, that copy is not optimistic — it is wrong, and it is wrong in a way that costs them. They choose the product partly on that basis, and every slow minute afterwards is measured against a promise we did not keep. Worse, the copy blamed their hardware for a limitation that was ours.

So we deleted it. Not softened, not qualified — removed, along with a set of performance figures that were quoted from an old version and no longer described the shipping product. The app no longer claims GPU speed anywhere, and it no longer explains slowness by pointing at your machine.

Why removal came before the fix

The obvious objection is that we should have fixed the attachment first and kept the claim. We disagree, for a reason worth stating plainly.

A claim is a promise made now, to someone deciding now. If the promise is untrue for some readers today, the fact that it might become true later does not help them — they will have already installed, waited, and drawn a conclusion about whether this software tells the truth. Trust is spent at the moment of the claim, not at the moment of the fix.

The fix continues as its own work. When GPU acceleration attaches reliably enough to be described honestly, we will describe it — with numbers measured on shipping builds, not aspirations.

What LocalSynapse actually promises about speed

We design for CPU. That is not a consolation prize; it is a set of deliberate properties:

• Search is available while indexing continues. Files become searchable by content as they are read, so you are not waiting for a finish line before the product is useful.
• Indexing runs under a ceiling. Each cycle gets a bounded budget instead of consuming whatever is free. Slightly slower in the best case, considerably more pleasant in the worst — which is the case people actually experience.
• Memory is handed back. The semantic model is released as soon as there is nothing left to index, rather than being held for the sake of an occasional instant search.
• Priority over completeness. The files most likely to matter are processed first, so useful results arrive long before everything is done.

If you are comparing local search tools, the useful question is not "does it support GPU?" — nearly everyone will say yes. It is "what does it do when the GPU is not available?", because for a meaningful share of real machines, that is the path you will actually be on.