This iPad Pro + Thunderbolt R9700 is running Qwen3.8-27B at 158.5 tps. Yes… an iPad. The trick 🎮 32GB Radeon AI PRO R9700 🔌 Thunderbolt eGPU enclosure 📱 iPad Pro running LemonSeed Studio 🧠 Qwen3.8-27B Q4 📚 131K context And the speed progression is kind of ridiculous 🐌 baseline → 31.2 tps ⚡ MTP3 → 108.7 tps 🚀 DFlash2 → 158.5 tps Prefill: 📥 4K → 1,636 tps 📥 32K → 1,395 tps LemonSeed actually runs the AMD GPU on iPadOS. It embeds the upstream Linux amdgpu + amdkfd driver as an iPad PCIDriverKit extension, then runs its inference engine on the R9700. Same engine for 🐧 Linux | 🍎 macOS | 📱 iPadOS Same R9700 on macOS reaches 159.4 tps, so the iPadOS result is basically right there with the Mac. ⚠️ Obviously, the iPad's own GPU isn't doing the 159 tps. The iPad is hosting a 32GB desktop AMD GPU over Thunderbolt. 🔗 Link in ALT.

Over a fast thunderbolt 5 link you can transfer tens of GB in 2 minutes, with a Python script, or in 30 minutes with scp... This is a mix of many inefficiencies, it's not encryption overhead. Would be cool to have more decent tooling: security is often an excuse for bad things.

MLX's implementation of RDMA (Remote Direct Memory Access) over Thunderbolt on macOS, can now be used as an independent library by anyone: It is the gem that powers Mac clusters for local AI, and is an order of magnitude faster than protocols over TCP.

I plugged a 100-gigabit datacenter network card into a MacBook Pro running Linux. ConnectX-5 Ex in a Thunderbolt dock on an M2 Max. The driver probed it, read its firmware and brought up both ports. Next: a cable, then a DGX Spark. Open source guide:

It’s so funny how the MCU post-Endgame wanted to adapt the same Avengers-less vacuum that created the Thunderbolts in the comics but since half of their heroes were still alive (unlike the Onslaught event) it just seems like nobody bothered to reassemble them.

Introducing UltraStudio Mini 12G models! Three new extremely portable Thunderbolt™ 4 capture and playback models with HDMI and dual 12G-SDI for 10-bit fill and key playback! Learn more at