WildID › GPU, VRAM and speed

GPU, VRAM and speed

WildID is built for strong NVIDIA graphics cards and for Macs with Apple Silicon. Every number on this page was measured on one graphics card, an RTX 4090, or on one Mac, a MacBook Pro with an M5 Pro. Nothing else has been tested, so read the notes about smaller graphics cards as estimates.

What it was tested on

One graphics card: an NVIDIA GeForce RTX 4090 with 24 GB of memory, in two computers. One runs Windows 11 with Lightroom Classic 15.4.1, the other runs Linux. No other card has been tried so far. The processor-only mode was run once, on the Linux computer.

One Mac: a MacBook Pro with an M5 Pro chip and 48 GB of memory, running macOS 26.5. No other Mac has been tried. Its figures come from the command line and from the engine fed the way the Lightroom plugin feeds it. The figures for the Mac

Speed on the RTX 4090

SituationPhotos per second
Engine as Lightroom uses it: 61-megapixel RAW files, crop on, place listabout 10
Command line, folder of RAW files, crop on27 to 32
Command line, folder of RAW files, no crop53 to 59
Command line, folder of JPEGs (1440 × 1800 pixels), crop on62 to 71
Command line, folder of JPEGs, no crop144 with BioCLIP 2, 98 with BioCLIP 2.5
Command line, crop and SpeciesNet second opinionabout 30
BioCLIP 2 unless stated. With RAW files and the crop, reading the files and finding the animal take most of the time. In Lightroom, writing keywords and collections into the catalog comes on top.

At 10 photos per second, a thousand photos take under two minutes of engine time. The first run of a session adds the time to start the engine and load the model, about ten seconds on the test computer.

Graphics memory per model

These figures are what the engine itself held while Lightroom sent it batches of 24 photos with the crop on. Add what Windows, Lightroom and your other programs already use, which was 1 to 2 GB on the test computers.

LoadedIdleWhile working
BioCLIP 24.8 GB6.8 GB
BioCLIP 2.57.7 GB9.7 GB
Both models12.3 GB14.1 GB

The engine gives the memory back when it stops. It stops by itself after 15 minutes without work, and you can stop it at once with WildID: server status… in Lightroom.

Smaller graphics cards

Nobody has run WildID on a smaller card yet. The figures above suggest the following, and these are estimates:

  • BioCLIP 2 with the crop needs about 7 GB free. A card with 12 GB should be comfortable. A card with 8 GB is tight, because the desktop and Lightroom use part of it.
  • BioCLIP 2.5 with the crop needs about 10 GB free. A card with 16 GB should be comfortable, one with 12 GB is tight.
  • Both models loaded need about 14 GB plus the desktop. The option is offered when the card reports 20 GB or more.

If the model does not fit into the graphics memory, the run can fail or become very slow. In that case use BioCLIP 2, switch off “Keep both models loaded” and close other programs that use the graphics card.

On a PC only NVIDIA cards are used. With an AMD or Intel graphics card WildID runs on the processor.

Why Lightroom shows 25.8 GB for a 24 GB card

The dialog reports graphics memory in decimal gigabytes, the same unit hard-disk makers use. A card sold as 24 GB shows as 25.8 GB and a card sold as 16 GB shows as 17.2 GB. The 20 GB rule for keeping both models loaded uses this reported value.

On a Mac with Apple Silicon

On a Mac, WildID uses the graphics processor that is part of the Apple chip. The figures below are what one Mac did: a MacBook Pro with an M5 Pro chip and 48 GB of memory. No other Mac has been tried.

SituationBioCLIP 2BioCLIP 2.5
Engine fed batches of 24 RAW files with the crop on, the way the Lightroom plugin feeds itabout 1510 to 11
Command line, folder of RAW files, crop on1611
Command line, folder of RAW files, no crop22 to 26not measured
Command line, folder of JPEGs (1440 × 1800 pixels), crop on2113 to 14
Command line, folder of JPEGs, no crop3317
Command line, crop and SpeciesNet second opinionabout 12not measured
Processor only, folder of JPEGs, no cropabout 8about 4
Photos per second on the M5 Pro. On the same command-line runs the RTX 4090 is two to six times as fast.

The first run of a session adds the time to start the engine and load the model, about four seconds for BioCLIP 2 on this Mac.

Memory on a Mac is shared between the system, your programs and the graphics processor, so WildID takes its share from the same memory Lightroom uses. These figures are what the engine held while it was fed batches of 24 photos with the crop on.

LoadedIdleWhile working
BioCLIP 29.4 GB11.7 GB
BioCLIP 2.512.9 GB13.4 GB
Both models19.2 GB20.1 GB

“Keep both models loaded” is offered on a Mac when it reports 30 GB or more as usable for the graphics processor. The tested Mac reports 40.2 GB of its 48 GB, and the Identify dialog shows that figure.

A Mac and a PC do not always give the same name

Without the crop, the tested Mac and the RTX 4090 gave the same name for every one of 863 JPEGs with both models and for every one of 157 RAW files. With the crop on, 37 of their 2,040 answers differed, about 2 % (1,020 photos, each identified with both models), and in 31 of those 37 the photo was flagged for review on both computers.

The reason is a small step before the animal detector. Each photo is reduced to the detector’s size, and the library that does this computes slightly different pixels on a Mac than on a PC: about a fifth of the values differ by 1 of 255. The detector then draws its box a pixel or two differently on about a third of the JPEGs, and where the model was undecided, another name comes out on top. Neither computer is wrong.

The Identify dialog has an option for this, “Exact detector resize (the same on a Mac and a PC)”, and the command line has --crop-exact. Both are off by default. With the option on both computers the detector received the same pixels on all 1,020 photos, and 5 of the 2,040 answers still differed, all 5 on photos flagged for review on both. Those 5 come from the PC’s default number precision: from the command line, with the PC set to the Mac’s precision (--precision fp32), no answer differed.

The option does not make the names more accurate, and turning it on changes about 2 % of them on either computer. It is there for when two computers have to agree.

Without a graphics card

WildID falls back to the processor by itself. It then needs about 10 GB of free main memory and takes seconds per photo. The one measurement on a PC so far: a single test photo took 6.9 seconds on a 32-core desktop processor, including the time to load the model. That is fine for a handful of photos and slow for a whole trip.

Disk space and downloads

PartSizeWhen
Windows installerabout 20 MBDownload
macOS downloadabout 23 MBDownload
Python and libraries, including PyTorch, on Windows and Linuxabout 5 GBInstallation
Python and libraries, including PyTorch, on a Macabout 1.2 GBInstallation
BioCLIP 24.5 GBSetup
BioCLIP 2.57.3 GBSetup, optional
MegaDetector crop stage52 MBSetup
SpeciesNet209 MBSetup, optional
Place listssmallWhenever you fetch one

Previous: ModelsNext: When a name is wrong