Jeff GeerlingModify

If you were 5 microseconds late today, blame it on NIST.

Their facility in Boulder Colorado just had its power cut for multiple days. After a backup generator failed, their main ensemble clock lost track of UTC, or Universal Time Coordinated.

But even if you used the NTP timing servers they run, they were never off by more than 5 microseconds.

5 μs might seem insignificant. But it is significant for scientists and universities who rely on NIST's more specialized timing signals.

But no, you don't need to panic. And yes, they have it under control now.

But I thought I'd go over what happened, what it means, and what we can learn from NIST's near-outage.

Video

This blog post is a lightly-edited transcript of my most recent YouTube video:

Ever since I got AMD, Intel, and Nvidia graphics cards to run on a Raspberry Pi, I had a nagging question:

What's the point?

The Raspberry Pi only has 1 lane of PCIe Gen 3 bandwidth available for a connection to an eGPU. That's not much. Especially considering a modern desktop has at least one slot with 16 lanes of PCIe Gen 5 bandwidth. That's 8 GT/s versus 512 GT/s. Not a fair fight.

Apple gave me access to this Mac Studio cluster to test RDMA over Thunderbolt, a new feature in macOS 26.2. The easiest way to test it is with Exo 1.0, an open source private AI clustering tool. RDMA lets the Macs all act like they have one giant pool of RAM, which speeds up things like massive AI models.

This little postage stamp is actually a full Raspberry Pi Zero 2, complete with eMMC storage and WiFi.

But you can't get one. Well, not unless you buy the CM0NANO development board from EDAtec, or you live in China.

This little guy doesn't have an HDMI port, Ethernet, or even USB. It's a special version of the 'Compute Module' line of boards. Little Raspberry Pi 'System on Modules' (SoMs), they're called.

Compute Modules are entire Linux computers about the size of a regular desktop CPU that you 'plug in' to another board, to give it life.

Now that Nvidia GPUs run on the Raspberry Pi, I've been putting all the ones I own through their paces.

Many people have an older Nvidia card (like a 3060) laying around from an upgrade. So could a Pi be suitable for GPU-accelerated video transcoding, either standalone for conversion, or running something like Jellyfin for video library management and streaming?

That's what I set out to do, and the first step, besides getting the drivers and CUDA going (see blog post linked above), was to find a way to get a repeatable benchmark going.

Inside this Framework 13 laptop is a special mainboard developed by DeepComputing in collaboration with Framework. It has an 8-core RISC-V processor, the ESWIN 7702X—not your typical AMD, Intel, or even Arm SoC. The full laptop version I tested costs $1119 and gets you about the performance of a Raspberry Pi.

A Pi 4—the one that came out in 2019.