Fitting 2x R9700 in one Strix Halo box via NVMe slots, with RCCL tuning
Pyrolistical · reddit · 2026-09-21
A Reddit user installed two R9700 GPUs in a GMKtec EVO-X2 (Strix Halo) — a machine with no oculink or PCIe slots — using NVMe-to-USB4 enclosures and NVMe-to-PCIe adapters in its two NVMe slots, both links running at PCIe 4.0 x4.
Debugging journey: llama.cpp's Vulkan backend botches tensor parallelism (slower than a single card); switching to the ROCm build improved decode but prompt processing ran at only 70% of single-card speed until building with -DGGMLHIPRCCL=ON to enable RCCL.
Final llama-bench results (Qwen3.8-27B-UD-Q4KXL): tensor parallel pp512 1069 tok/s, tg128 37 tok/s, vs single card 1070 / 28 — parity on prefill, 32% faster decode.
More from Infra
- Laya, an open-source local AI, is being built into Omarchy M to run on Apple GPUs and ANE — Scobleizer · 2026-09-21
- Facebook and Instagram down for thousands of users in ongoing outage — Polymarket · 2026-09-21
- Agent builders say provider KV caching black boxes block swarm and long-run agents — Small_Luck8177 · 2026-09-21
- Laya ported to MLX: M3 Max runs local agent 60 decisions/sec, 50x faster — sven_ai · 2026-09-21
- US Data Center and Info-Processing Hardware Spending Now Exceeds Housing Investment — rohanpaul_ai · 2026-09-21
- Intel's BITCOS compresses ternary LLMs to 1.485 bits per weight, boosting decode up to 27% — burny_tech · 2026-09-21