Minimum requirement: an NVIDIA GPU
Required
BallTrack AI runs a YOLO neural network on every single
frame to find the ball and the players. That detection
runs on CUDA, so an NVIDIA GPU is required, not
optional — without one the app has nothing to run on.
There is no usable fallback: a CPU / Apple Silicon code path
exists, but it is orders of magnitude too slow to
process a full match, and it is not offered as an option.
VRAM is the setting you buy. The app lets you pick
how many detection workers to run — a
1–5 choice that starts at 1, and the main speed control you
have. Each worker
is a separate process holding its own copy of the model on the GPU,
about 2 GB each, so VRAM caps how high you can
set it. Going from one worker to three is a large win; past about
four the GPU is saturated and more workers stop helping. That makes
the useful target enough memory for 3–4 workers,
not the largest number you can afford.
-
10–11 GB VRAM
3 workers — about 8% off full speed
-
12 GB VRAM
Recommended minimum — 4 workers, full speed
-
16 GB VRAM
Sweet spot — all 5 workers, or a heavier model at fewer
-
24 GB VRAM
The only tier actually validated — but no faster than 16 GB
12 GB is the practical floor. An RTX 3060
12 GB is the cheapest card we would recommend: it reaches four
detection workers, which is where the gains flatten, and it has a
working NVENC encoder. 10–11 GB cards (RTX 3080
10 GB, RTX 2080 Ti) do run, but top out at three
workers. 8 GB is below the floor — one or
two workers, roughly a quarter of the pace given up. Cards older
than the RTX 20 series are not supported at all. Picking a
heavier model raises the cost per worker (the largest is about
5 GB each), so model choice and worker count trade against one
another. See the
GPU performance table for
per-card figures and the worker ladder.
Honest caveat: those tiers are arithmetic, and only
the top one has been measured. The one validated setup is five workers
plus the parent process using about 11 GB on a 24 GB card.
12–16 GB is comfortable on paper and is what we recommend,
but only 24 GB has actually been run.
FullField Stitch above has no such
requirement — it stitches on any OpenCL-capable GPU, NVIDIA or
not, and its planned Apple Silicon build needs no NVIDIA
hardware at all. The hard NVIDIA requirement applies to BallTrack
AI only.