Pinned post
How does an AI-designed PCB compare on EMI with the chip vendor's own reference design? Drum roll, please, again!
A year ago, we manufactured our Project Speedrun, the first computer designed by AI. Quilter's physics-driven AI placed and routed both boards of an NXP i.MX 8M Mini Linux system, and engineers spent 38.5 hours on cleanup.
EMI was the first question people asked about Speedrun and the most common criticism. We're glad to answer it now with third-party measurements.
None other then Kenneth Wyatt, the EMC Doctor, measured Project Speedrun against a board built from NXP's i.MX 8M Mini reference design files, in a TEM cell in his lab.
✅ Through much of 250 to 650 MHz, the Speedrun trace sat below the NXP-design trace, with a gap of about 12 dB in part of that range by Ken's estimate
✅ Broadband DC-DC converter noise from 10 to 250 MHz was about 5 dB lower
✅ The likely Ethernet clock harmonics were lower too, with the same Ethernet PHY on both boards
These are bench pre-compliance measurements at the TEM cell output, so they don't predict margins at a 3 m or 10 m test distance. The write-up covers the setup and the limits.
quilter.ai/blog/ai-design…
A year ago, we manufactured our Project Speedrun, the first computer designed by AI. Quilter's physics-driven AI placed and routed both boards of an NXP i.MX 8M Mini Linux system, and engineers spent 38.5 hours on cleanup.
EMI was the first question people asked about Speedrun and the most common criticism. We're glad to answer it now with third-party measurements.
None other then Kenneth Wyatt, the EMC Doctor, measured Project Speedrun against a board built from NXP's i.MX 8M Mini reference design files, in a TEM cell in his lab.
✅ Through much of 250 to 650 MHz, the Speedrun trace sat below the NXP-design trace, with a gap of about 12 dB in part of that range by Ken's estimate
✅ Broadband DC-DC converter noise from 10 to 250 MHz was about 5 dB lower
✅ The likely Ethernet clock harmonics were lower too, with the same Ethernet PHY on both boards
These are bench pre-compliance measurements at the TEM cell output, so they don't predict margins at a 3 m or 10 m test distance. The write-up covers the setup and the limits.
quilter.ai/blog/ai-design…