Aurora gaming PC¶
Status¶
Active — campus dorm Windows gaming PC (not on the home LAN). It also hosts dorm Home Assistant and, as of 2026-10-06, the primary Ollama / Open WebUI stack. Hardware was filed 2026-09-24. Daily Afterburner as of 2026-09-30 is still unlocked +325/+1250, PL 120%, Curve Editor unlocked. 9600X Curve Optimizer all-cores −15 is applied. EXPO was re-enabled the same day so DDR5-6000 should apply. 2026-10-01: Meta Quest Steam Link over Aurora's Windows hotspot is an active constraint / open loop (not a solved deploy). 2026-10-27: home 10.0.0.x is reachable only through Leviathon over Tailscale. Tailscale DNS (CorpDNS) stays off on this PC. 2026-10-28: Tdarr on Seamoth uses this PC’s RTX 5070 as a remote transcoding worker.
Description¶
Aurora is the always-on Windows gaming PC on the campus dorm network (not the home LAN). It reaches home 10.0.0.x only through Leviathon over Tailscale. Dorm HA, Zigbee, presence lighting, and SignalRGB stay on projects/Self-Hosted-Services#dorm-home-assistant-aurora. As of 2026-10-28 it is also the remote transcode worker for Tdarr on Seamoth. This page is the hardware, overclock, local game-streaming, local AI, and Tdarr-worker profile.
Hardware¶
Confirmed 2026-09-23–24:
| Part | Spec |
|---|---|
| CPU | AMD Ryzen 5 9600X (user believes PBO is already enabled) |
| Board | ASRock B650M Pro RS WiFi |
| RAM | 32 GB DDR5-6000 CL30 (EXPO/XMP) — keep EXPO/XMP on |
| GPU | NVIDIA RTX 5070 12 GB |
| PSU | Corsair RM750e — adequate for this build; no PSU upgrade needed |
| Cooler | Thermalright Frozen Notte 360 AIO |
EXPO / RAM (2026-09-30)¶
EXPO was re-enabled on the ASRock B650M Pro RS WiFi so DDR5-6000 should apply. Earlier the same day the kit was running at 4800 with EXPO/XMP off. That, with nearly full 32 GB and heavy paging, contributed to perceived slowness. After the EXPO and Curve Optimizer changes, RAM usage looked fine (~70%); paging was no longer the story. DRAM speed was not re-verified at 6000 MT/s in Task Manager.
RTX 5070 Afterburner (daily profile)¶
Set 2026-09-25; still the daily profile on 2026-09-30. Unlocked offsets; Curve Editor left alone (no locked yellow voltage line).
| Core / memory | +325 / +1250 |
| Power limit | 120% |
| Curve Editor | Left alone (unlocked; no locked yellow voltage line) |
| Core Voltage | +100% if unlocked (or +0 if locked) |
Try +325 / +1250 first. If that is unstable, drop core to +300 or lower memory. Retest a full Dead Space Remake session for stability before raising core again.
Dead Space Remake hang (2026-09-25)¶
At Afterburner +350 core / +1250 mem, Dead Space Remake hung with DXGI_ERROR_DEVICE_HUNG (driver 61692). +350 is unstable in that game.
Ladder (history)¶
Dialed 2026-09-23: +325/+1250 → +350/+1250 → +375/+1250 (gains flattened), then +350/+1350 tested. At the then-daily +350/+1250 profile, GPU-Z load was ~300 W / ~3.1 GHz / ~1.05 V and Steel Nomad ~5,830–5,843 (example result: 3DMark 164275888). Voltage-curve lock and undervolt-curve experiments collapsed clocks and score. Curve Editor was wiped; the unlocked +350/+1250 profile was recovered. On 2026-09-25, +350/+1250 was marked unstable in Dead Space Remake, and the daily target moved back to +325/+1250.
Dead Space Remake loads (2026-09-30)¶
Still on the daily Afterburner profile (+325/+1250, PL 120%).
Door/area load stall: loads stall about 20 seconds with disk at 100% while Dead Space itself only reads about 20–30 MB/s. The drive is busy but barely moving data. Closing other apps will not fix it. A drive upgrade is the actual fix. Waiting on a cheaper drive upgrade.
Half-load glitch (separate): half-loaded scenes / hung door loads even when the drive is idle (low active time). That is not the disk stall. EA repair does not clear a hung load. Earlier the same day, half-loads were also tied to the Afterburner overclock causing DXGI device hangs (consistent with the 2026-09-25 +350 DXGI_ERROR_DEVICE_HUNG note). Daily stays +325/+1250.
9600X Curve Optimizer (applied)¶
2026-09-30: all-cores −15 is applied (no longer deferred). It posted and held about 5.33 GHz at ~80% load (near the 5.4 GHz boost). This is a light-load observation so far. Drop to −10 if Dead Space hard-crashes or WHEA errors appear. Tested separately from the GPU Afterburner OC. EXPO/XMP stays on (re-enabled the same day; see above).
Steam Link / Meta Quest¶
Status: active constraint / open loop — not a solved deploy. Recorded 2026-10-01 (CS Lab / agent memory).
The user has a Meta Quest and streams PC games from Aurora over Steam Link.
The dorm has no Ethernet drop and no spare router, so Quest Steam Link cannot use a wired PC or a second router. It has to go through Aurora's Windows Wi-Fi hotspot.
Steam Link from Aurora to the Meta Quest was very poor while the headset used that hotspot: about 631 ms ping and about 78% packet loss. The hotspot is the bottleneck, not an internet hairpin.
Preferred setup (not available in the dorm): Ethernet on Aurora shared as the hotspot, or the Quest on a real router on the same LAN so the PC shows as a local Steam Link computer.
Interim plan (not done): ease with 5 GHz; later a USB Wi-Fi adapter so one radio is not doing both internet and hotspot.
Local AI stack (2026-10-06)¶
Aurora is the primary Ollama host. The RTX 5070 has about 12 GB VRAM.
Open WebUI runs in Docker Compose on WSL. It is reachable over Tailscale at http://aurora. Tailscale Serve publishes :80 to local 3000. Home Assistant stays on https :443.
| Daily default model | qwen3:14b-chat |
| Context | Raised to 6144 |
| Built-in tools | Disabled — they cost about 4,200 tokens |
| Web search | SearXNG, in the same compose |
| DB backups | Under the compose folder’s backups/ subfolder |
Custom filters on Open WebUI:
- Context meter — shows usage as a percent of the model’s limit.
- Compaction — a same-model summarization filter that injects a short history summary as a system message.
qwen3-vl has a dedicated model entry plus a system prompt so images work in fresh chats.
Snowfox (secondary Ollama)¶
Snowfox (MacBook Pro M5 Pro, 24 GB unified memory) runs a secondary Ollama over Tailscale only while in use. Models:
| Model | Notes |
|---|---|
| qwen3:30b | MoE |
| qwen3:14b-long | 16k context |
| qwen3-vl:8b | Vision |
They show in the picker as snowfox/....
Tailscale on Aurora¶
Aurora sits on the campus dorm network, not the home LAN, so it reaches home 10.0.0.x only through Leviathon over Tailscale. Tailscale DNS (CorpDNS) must stay off on Aurora. Those tailnet DNS servers are only reachable that way. A Tailscale hang killed all internet. Tailscale auto-start races also caused drops.
Tdarr remote worker (2026-10-28)¶
Aurora’s RTX 5070 is the remote transcoding worker for Tdarr on Seamoth (http://10.0.0.20:8265). The Tdarr server stays on Seamoth. The libraries hold new files for 31 days and skips hardlinked (still-seeding) files. Movies rules re-encode H.265 only when the file is above about 8 Mbps at 1080p. The goal is to save space on bloated H.265, not to skip H.265. The homelab media library is close to 1 TB and server storage is getting tight. projects/Media-Automation-Stack-Seamoth#tdarr-2026-10-28.
Considering (not done)¶
- A future 9800X3D for CPU-bound games. As of 2026-09-24, SSD interest was mainly capacity. As of 2026-09-30, a cheaper drive upgrade is the planned fix for Dead Space door/area load stalls (waiting on one).
- Open question only: whether the ASRock B650M Pro RS WiFi still carries a 9800X3D frying risk as of late 2026. No verdict is recorded.
Known issues¶
- 2026-09-25: +350/+1250 hung Dead Space Remake with DXGI_ERROR_DEVICE_HUNG (driver 61692). Daily target is +325/+1250. Retest a full session before raising core.
- 2026-09-30: Dead Space Remake door/area loads stall ~20 seconds with disk at 100% while the game only reads ~20–30 MB/s. Closing other apps will not fix it. Waiting on a cheaper drive upgrade.
- 2026-09-30: separate half-loaded scenes / hung door loads when the drive is idle (low active time). EA repair does not clear a hung load. Earlier that day, half-loads were also tied to Afterburner OC DXGI device hangs.
- Curve Optimizer all-cores −15 is applied. Light-load only so far (~5.33 GHz at ~80% load). Drop to −10 if Dead Space hard-crashes or WHEA errors appear.
- A voltage-curve lock on the 5070 collapsed clocks; the daily profile stays unlocked offsets (Curve Editor left alone).
- 9800X3D on this board: frying risk is an open question.
- 2026-10-01: Meta Quest Steam Link from Aurora is very poor on Aurora's Windows Wi-Fi hotspot (~631 ms ping, ~78% packet loss). Dorm has no Ethernet drop and no spare router, so a wired PC or a second router is not available. The hotspot is the bottleneck, not an internet hairpin. Open loop — not solved. Interim plan: 5 GHz, then a USB Wi-Fi adapter so one radio is not doing both internet and hotspot.
- 2026-10-06: Tailscale DNS (CorpDNS) must stay off on Aurora. Tailnet DNS
10.0.0.20/10.0.0.21is only reachable via Leviathon; a Tailscale hang killed all internet. Tailscale auto-start races also caused drops. - 2026-10-27: Aurora is on the campus dorm network, not the home LAN. Home
10.0.0.xis reachable only through Leviathon over Tailscale. CorpDNS stays off.
Raw sources¶
- 2026-09-23–24 CS Lab / agent-memory session (hardware profile, Afterburner OC ladder, Curve Optimizer plan).
- 2026-09-25 CS Lab / agent-memory (Dead Space Remake DXGI_ERROR_DEVICE_HUNG at +350/+1250, driver 61692; daily target +325/+1250, PL 120%, Core Voltage guidance; Curve Optimizer still deferred at that date).
- 2026-09-30 CS Lab / agent-memory (EXPO re-enabled on the B650M Pro RS WiFi after running DDR5-4800 with EXPO/XMP off; nearly full 32 GB / heavy paging; after EXPO + CO, RAM ~70% and paging was no longer the story; DRAM not re-verified at 6000 MT/s in Task Manager. 9600X Curve Optimizer all-cores −15 applied, ~5.33 GHz at ~80% load, light-load observation, tested separately from the GPU OC. Dead Space Remake door/area stalls ~20 s at 100% disk / ~20–30 MB/s reads — waiting on a cheaper drive upgrade. Separate half-loaded scenes / hung door loads when the drive is idle; EA repair does not clear a hung load).
- 2026-10-01 CS Lab / agent-memory (Meta Quest streams PC games from Aurora over Steam Link. Dorm has no Ethernet drop and no spare router, so the headset has to use Aurora's Windows Wi-Fi hotspot. Measured ~631 ms ping and ~78% packet loss; the hotspot is the bottleneck, not an internet hairpin. Preferred but unavailable: Ethernet on Aurora shared as the hotspot, or the Quest on a real same-LAN router so the PC shows as a local Steam Link computer. Interim plan, not done: ease with 5 GHz; later a USB Wi-Fi adapter so one radio is not doing both internet and hotspot).
- 2026-10-06 weekday ingest (Aurora primary Ollama; Open WebUI in Docker Compose on WSL at http://aurora, Tailscale Serve :80 to 3000, Home Assistant stays https :443; daily qwen3:14b-chat, context 6144, built-in tools disabled at ~4,200 tokens; SearXNG in the same compose; context-meter and same-model summarization compaction filters; qwen3-vl dedicated model entry plus system prompt. Snowfox secondary Ollama only while in use: qwen3:30b MoE, qwen3:14b-long 16k, qwen3-vl:8b, picker prefix
snowfox/. Open WebUI DB backups under the compose folder’sbackups/. CorpDNS must stay off; tailnet DNS10.0.0.20/10.0.0.21only via Leviathon; Tailscale hang killed internet; auto-start races caused drops). - 2026-10-28 weekday ingest (facts as of 2026-10-27): Aurora sits on the campus dorm network, not the home LAN, and reaches home
10.0.0.xonly through Leviathon over Tailscale. CorpDNS stays off. - 2026-10-29 weekday ingest (facts as of 2026-10-28): Aurora’s RTX 5070 is the remote transcoding worker for Tdarr on Seamoth (
http://10.0.0.20:8265). The libraries hold new files for 31 days and skips hardlinked (still-seeding) files. Movies rules re-encode H.265 only above about 8 Mbps at 1080p (save space on bloated H.265, not skip H.265). Media library close to 1 TB; server storage getting tight.
Related work¶
- projects/Self-Hosted-Services#dorm-home-assistant-aurora — dorm HA on this same PC.
- projects/Media-Automation-Stack-Seamoth#tdarr-2026-10-28 — Tdarr server on Seamoth; this PC is the transcode worker.