AMD
Maker · 6 entries
Everything from AMD
| Name | What to know | Price or memory | |
|---|---|---|---|
| Eight R9700 cards at home run local Qwen and GLM models for debugging, kernel and driver work, per one user | |||
| MI355X running vLLM with open FlyDSL kernels beats Blackwell on MiniMax M3.1, a more than 3x improvement | |||
| 16B MoE (2.8B active) trained from scratch on AMD Instinct GPUs with ROCm; checkpoints, recipes and code published | ~32 GB | ||
| Overlaps expert-parallel GPU data exchange with compute; AMD reports 12.7% shorter pre-training and up to 39.2% lower time to first token | |||
| Learned gate after multi-head latent attention that turns down weak signals, credited for strong results with few active parameters | |||
| Single APU (Radeon 8060S, gfx1151, ~125GB visible unified memory) is enough to run a 284B MoE model locally | |||