Render on another Mac
- Render › Compute On moves the whole window to another Mac of your network that runs Siliconed: a
MacBook in the living room renders on the Mac mini in the next room. The queue, the progress, the
preview, the history, the models, the library, the licenses and the capacity map are the other
Mac's; the form, your selection, exports and the clipboard stay on yours. One menu brings you back.
Never two views at once: an “On:” menu beside Chain says where you compute, and the rack wears a
blue edge while connected. - Pairing by a six-digit code shown on both screens and accepted on the lending Mac; then TLS 1.2
with a key per paired Mac (kept in the Keychain), nothing in clear on the network. The lending Mac
turns lending on in Settings › Other Macs. - Everything you do here works there: renders, edits with your references, variations, grids and
explorations, enlarging and cropping, named subjects, undo; installing a model or importing a LoRA
on the lending Mac (removing one waits for a click on that Mac); measuring it. Each Mac sees only
its own jobs and images.silicontrolfollows Compute On.
The capacity map, rebuilt
- The map now chooses a route for each computation of a render, not one switch for all: each
matrix product at each height, each attention, on the GPU alone, beside the AMX, beside the Neural
Engine or the three, with the weights widened or read as published. On a Mac where the Neural
Engine wins, it is used where it wins. - The map knows no model: it measures a fixed grid of products and attentions that spans the
shapes every model is made of, the same on every Mac, whatever is installed. Installing, removing
or adding a model never asks for a new map, and a Mac with no model measures it whole. - Any format, any shape: a computation the map did not measure — a 1024×1536 image, an edit —
takes the route of the nearest one it measured, and only a shape far from all of them stays on the
GPU alone. Before, everything outside the measured formats ran on the GPU alone. - The slowest Mac fits its minute: a route not yet timed at one shape is judged from its timings
at the others, and each second of the map goes where it settles the most. On a MacBook Neo the map
used to run out of time before it reached the Neural Engine; it now finds it at every product it
measures, 1.7 to 2.7 times as fast as the GPU alone. - Measured the way a render runs: each route is timed against the simplest in alternation, so a
Mac that heats up or runs on battery compares like with like; the routes are chosen at the warm
regime a render of a few minutes runs at, with one witness per engine. A product is timed once on
the engines and decided for each form its weights can be published in. About a minute. - Green when the measure is complete and right, not when the Mac was at rest: heat, the battery
and Low Power Mode are written beside the map, no longer a reason for it to be orange. - Under the memory budget: the map's own need is computed before its first product, as a
render's; it now stays under B on an 8 GB Mac (it peaked at 4.9 GB before). - Each render checks the map: what each computation cost against what the map predicted; a gap
corrects the map at the next launch, or measures it again.
Also
- The app in Japanese and Simplified Chinese.
- Lighter: 19 MB installed instead of 25, the binaries without their debugging symbols; the same code.
- Settings in three tabs: General, Other Macs, Advanced.
- One schedule: the ComfyUI sigma schedule is gone; the official one is the only one.