🧵 Mega thread on RF, SDR, ham radio, and signal hacking: I've been writing educational posts and threads on these topics. To help finding them easier, I will put all the links here. And I will link the new threads to the bottom of this meta thread every time I write one. 0/n

Thermal image of a Mac mini under heavy load. You can see the hot airflow.

My Mac Mini's GPU temperature at 90C running Qwen3.8 27B 24/7.

ISDT sent me their new banana output cable to test. I already prefer it to the original alligator-clip version. It protrudes much less, and saves desk space. Second pic shows them side by side.

People running local LLMs for a team: what does your cluster look like? I'm experimenting with 4 DGX Sparks, with 2 M5 Ultra Macs planned next. Has anyone scaled this kind of hardware into a larger cluster serving dozens of concurrent users?

DGX Sparks are cooking 🔥 GLM-5.3-Flash on TensorFold, mid-benchmark:

Initial results are very promising! GLM-5.3-Flash on 2x DGX Spark with TensorFold: 85 tok/s single stream, 174 tok/s total across 8 streams (code)

For people who will connect two Sparks: if your direct connection doesn't reach close to 180-190 Gbps, then do this: update the Linux and firmware using apt and fwupdmgr on both machines and then reboot them. That fixed my issue (initially it was 26 Gbps, now it's 185 Gbps)

What are you guys working on this weekend? I'm setting up and benchmarking GLM-5.3-Flash on 2 of my 4 DGX Sparks (and probably DeepSeek V4.1 Flash on the other pair). Once my 400G switch arrives, I'll run one model across all 4.

DGX Spark consumes ~25 W at idle (no model loaded). It uses USB-C Power Delivery 3.1 EPR (Extended Power Range), which supports up to 240 W (48 V, 5 A).

Amazon Prime just temporarily lowered its membership price in Germany due to a court ruling. In an economy where prices seem to only go up, this is an interesting one.