- Rename host alias matt-strix -> strix-halo (removes third-party name)
- Move host URLs to env-var lookup (OLLAMA_*_URL), drop hardcoded IPs
from harness source. Defaults: steel141 keeps localhost; pve197 and
strix-halo require their env var to be set before use.
- Update doc: remove the Tailscale IP and LAN-IP references, describe
access paths without specific addresses.
- Rename runs/matt-strix -> runs/strix-halo and patch the host field
in each JSON.
Harness still functional for the original author (set the env vars)
and safe to share without leaking routable addresses.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Cross-host Gemma 4 throughput comparison across three architectures.
Harness at scripts/gpu-bakeoff/; writeup at
docs/reference/gpu-bakeoff-2026-04-20.md.
Key findings:
- RTX 3090 Ti wins decode decisively (128 tok/s on gemma4:26b MoE Q4,
~4.7× faster than gemma4:31b dense on the same card).
- AMD Strix Halo iGPU lands at ~42% of 3090 Ti decode on ~25% of the
memory bandwidth — good SIMD utilization, especially for MoE.
- V100 numbers are DEGRADED: CT 167 ai-visualizer SDXL consumes 31/32
GB of its VRAM, forcing Gemma 4 models 95% onto CPU. Isolated V100
run requires SDXL eviction — left as follow-up.
- MoE vs dense is the dominant latency factor across all GPUs: ~4 B
active params of gemma4:26b beats 31.3 B active of gemma4:31b by
the same ratio (~4.7×) on every card tested.
Methodology: 1 warmup + 3 measurement runs per (host × model ×
prompt-length), Ollama's canonical timing fields, temp=0 greedy,
num_predict=256. All three Ollama servers accessed via HTTP (Strix
via Tailscale).
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Mortdecai