Silicon · Chapter 3 of 7
Sensor to compute
Eleven sensor streams fuse in 8.6 ms, leaving 74% of the frame budget spare.
DeepGrid Semi's in-cab compute part accepts seven RGB cameras, two thermal cameras and two radars, eleven configured inputs, and outputs a single fused AD2 perception stream. The reusable product is the compute and software platform; the sensor installation is one application of it, which matters commercially because the same semiconductor can serve multiple truck configurations.
- 11 inputsconfigured sensor channels
- 7 RGB camerascolour camera count
- 2 thermal camerasthermal camera count
- 8.6 msframe budget used (design target)
- 33.3 mstotal frame budget
- 74% headroomremaining frame budget (estimate)
- 7RGB camerasroad, sides, dashcam, driver, rear
- 2thermal camerasleft and right towers
- 24D radar77 GHz, front and rear
8.6 ms is a design target, not a measured production latency.
The animation opens with 11 sensor channels feeding into a single unit and one perception result coming out. That simplicity is the core design promise.
Seven RGB cameras at 8 megapixels over GMSL2 cover road left, road right, 120-degree sides, dash, driver monitoring, and rear. Two thermal cameras add coverage beyond visible light.
Front and rear 4D radar at 77 GHz join the camera array, giving the system active sensing alongside passive imaging.
The animation shows all 9 cameras and 2 radars converging on the processor and producing a single AD2 perception output in real time. This is a simulated demonstration of the design target, not measured silicon.
A progress bar shows the design target: the full perception workload consumes 8.6 milliseconds out of a 33.3-millisecond frame budget, figures from simulation on an FPGA-based SoC 1.x generation.
With 74% of every frame left unused as a design target, one tapeout can address 15 products rather than requiring 15 separate development programs.
Inspect full-size figure ↗Technical detail
The seven RGB cameras are assigned distinct roles: road left and right, side left and right, a forward dashcam, a driver-monitoring camera and a rear camera. Each mirror tower combines a road-facing camera, a 120-degree side camera and a thermal camera, so the same quarter of the truck is observed in three different sensing modalities simultaneously. Front and rear 4D 77 GHz radars contribute range and relative-motion data at both ends of the vehicle, where cameras are weakest in rain and glare. The Information Memorandum (June 2026) states the design target is to fuse all eleven channels in 8.6 ms of the 33.3 ms frame, described as 74% headroom; the architecture illustration is labelled 'illustrative sensor-fusion timing, not a measured production latency.' The chip is not yet fabricated in 28 nm, so these figures require workload and implementation validation; calibration, occlusion, mounting tolerance and the final sensor bill of materials also remain to be confirmed.
Portfolio deck, slide 24: Eleven inputsPortfolio deck, slide 25: The compute box, runningPortfolio deck, slide 26: Each sensor's jobInformation Memorandum (June 2026), page 10: The compute box: 11 sensors in, AD2 out