What Would UxHw® Save by Collapsing Your Ensemble Into One Pass?
A deep ensemble runs several models to find out how uncertain one answer is. Computing on distributions returns the spread from a single pass. Enter your member count, your host processor, and the module.
Models in the Ensemble
5
How many members you run to get a calibrated answer, or how many forward passes MC dropout takes. The published benchmark is five branches, so five is the like-for-like setting.
Inferences Per Device Per Day
8,640
One every ten seconds is 8,640 a day. This drives energy rather than cost, because the device is already bought.
Fleet Size20,000
Devices in the field. On a new design this is the annual build.
Signaloid Module
The fidelity ceiling matters more than price here. The published figure is measured at η = 32, which C0-microSD cannot reach, so the $49 part is the wrong part for this workflow.
Host Processor Cost
What the processor running the ensemble costs you per unit. No default: a third-party part price is yours to supply.
Host Processor Power
Power drawn while running all members. No default for the same reason.
Battery Capacity
Used only to express the power difference as runtime, which is the unit an embedded engineer budgets in.
C0-microSD
$49, shipping today from Mouser. 12 MHz, 128 KiB, 17.1 mW average, up to η = 16.
$49, shipping today from Mouser. 12 MHz, 128 KiB, 17.1 mW average, up to η = 16.
C0-microSD+
$99, engineering samples available today, contact sales@signaloid.com. At least 60 MHz, 320 KiB, up to η = 64.
$99, engineering samples available today, contact sales@signaloid.com. At least 60 MHz, 320 KiB, up to η = 64.
C0-SD
$199, engineering samples available today, contact sales@signaloid.com. 90 MHz, 448 KiB plus 128 MiB PSRAM, up to η = 128.
$199, engineering samples available today, contact sales@signaloid.com. 90 MHz, 448 KiB plus 128 MiB PSRAM, up to η = 128.
Benefits