desktop: full from-scratch AM5 build (3090 + 9800X3D), gift old tower intact #189
Labels
No labels
burndown-2026-06
coherence-core
consult
headless
interactive
P0
P1
P2
P3
P4
No milestone
No project
No assignees
1 participant
Notifications
Due date
No due date set.
Dependencies
No dependencies set.
Reference
coilyco-flight-deck/infrastructure#189
Loading…
Add table
Add a link
Reference in a new issue
No description provided.
Delete branch "%!s()"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
Goal
Full from-scratch desktop build for Kai, with two missions: run Qwen / local LLM (24GB VRAM) and fix the Factorio CPU bottleneck (late-game UPS). The current 2018-vintage platform (i7-8700 / Z390 / DDR4 / RTX 2080) has aged out as a whole, so this supersedes the drop-in-3090 plan in #85.
The old tower stays 100% intact and gets gifted to Kai's partner as a complete, working 1080p/1440p gaming PC. Nothing is cannibalized.
Supersedes
Parts already owned (carry into new build)
Parts to buy (Kai is doing diligence on each)
Estimated new spend: ~$2,650-2,800.
Open decisions
Acceptance
nvidia-smishows 24GB.desktop-tower.mdupdated (or a new machine doc created for the from-scratch build).Related
Spec update - GPU changed to 3090 Ti, sourcing reality, market context
Walked the full build component-by-component. Most of the spec firmed up; two things shifted hard due to the mid-2026 market.
GPU: RTX 3090 -> RTX 3090 Ti (sourcing forced it)
The plain 3090 turned out to be structurally unavailable from trusted retailers:
Decision: RTX 3090 Ti, ~$1,784, new, Sold-by-Amazon (or Newegg refurb FE fallback). Keeps CUDA (smoothest Qwen path on Windows) and 24GB, available now. Rejected alternatives: RX 7900 XTX (~$800, new, in stock, 24GB - rejected for AMD/ROCm friction, Kai wants CUDA simplicity), keep-hunting-a-3090 (rejected - Kai wants it now).
PSU pairing is actually ideal: the 3090 Ti FE's 450W + transient spikes are exactly what the already-owned RM850x (ATX 3.1, native 12V-2x6) is built for. Fits the O11 EVO XL easily.
RAM: price is real, not an error
64GB (2x32) DDR5-6000 CL30 is ~$1,160, not the ~$200 I first assumed. Confirmed: 2026 AI-driven DRAM shortage has pushed DDR5 up ~4x (32GB kit ~$90 -> ~$529), no normalization expected until 2027+. Corsair Vengeance CMK64GX5M2B6000Z30 is the chosen kit. (Open: Kai may reconsider 64GB vs 32GB now that the gap is ~$630, not ~$100 - but the X3D 2-DIMM constraint makes "upgrade later" a buy-twice trap, so leaning keep 64GB.)
Final spec (built around already-owned RM850x + SN7100)
$198) + PWM hub ($10). (Open: confirm T30-140 vs T30-120 - EVO XL supports 140mm, more airflow; and buy from a reputable seller, not a random marketplace store - counterfeit risk.)Rough new spend: ~$4,300 (RAM and the 3090 Ti are the two big line items, both market-inflated).
Setup gotchas captured
Still open (Kai's calls)
ORDERED - build placed, one loose end on the GPU payment
Both orders placed.
Newegg order - placed, clean ($3,053.33 all-in)
Item subtotal $2,778.92 + shipping $11.62 - $20 promo + tax $282.79 = $3,053.33. Verified line-by-line, no protection-plan padding. Contents:
Amazon order - placed, but charge NOT posting (LOOSE END)
All-in total: ~$4,750 with tax (Newegg $3,053 + Amazon ~$1,693)
Owned, carrying into build (not purchased): Corsair RM850x (850W, confirmed) + WD_Black SN7100 2TB NVMe.
Remaining acceptance work (post-arrival)
Backlog burndown 2026-06-17: closing low-priority (P3/P4) to bring the open count to a manageable level. Nothing lost — reopen if this resurfaces. Batch tag:
burndown-2026-06.