[phobos665]

The NV2A ack thread polls the GPU register words a title writes as plain memory. Nothing tells it when one changes, so while a title is busy it yields in a loop, which costs a fifth to a third of a host core for the whole run. That is power and thermal headroom the guest's own core needs on a laptop or a phone.

This turns on the busy backoff that was already there: RECOMP_NV2A_ACK_BUSY_US now defaults to 50 µs everywhere except Windows. After eight passes in a row that find nothing, the thread sleeps 50 µs between passes. Any change it finds puts it straight back to yielding.

It was left off because nobody had measured the titles that pace on the GPU fence. OutRun 2 is one of them, and it shows no cost.

Measurements

Apple M4, macOS. Runs alternate between 0 and 50 so machine load lands on both sides equally. Each run was muted and in the background.

Title	Ack thread, share of a core	Frame rate
OutRun 2 (fence-paced)	8% → 1.7%	60 fps either way, same frame times with the cap on and off (2 runs each)
TimeSplitters 2	20–28% → 2–3%	60 capped either way; uncapped 119 → 117 fps
Constantine	25–35% → 2%	60 capped either way; uncapped 94 → 92 fps

The only cost is 1.5–3% of the uncapped top end in TimeSplitters 2 and Constantine, from one run each per setting. Nothing is measurable at 60.

Scope

• Windows is unchanged. The default there stays 0 until someone measures it on Windows, so this needs no Windows regression.
• Not measured: Burnout 2, the other fence-paced title.
• To revert: RECOMP_NV2A_ACK_BUSY_US=0 restores the old loop on any host. RECOMP_NV2A_ACK_STATS=1 prints the thread's CPU use and how often a change was found right after a sleep.