Hardware explorer · phongvu.vn · prices 2026-06-16
Vietnam builds for a LAN coding-agent team
An interactive companion to
“Building a LAN coding-agent team before the API door closes.”
Toggle the tier, platform, and whether you reuse parts already in podmaster . Prices are live phongvu.vn listings (VND); USD uses today’s rate 1 USD ≈ 26,300 ₫ . Items marked est are estimates to confirm.
Target model band 12B–27B coding models. Reference: Gemma 4 12B Coder Q4_K_M (~7 GB) .
Speed target 50+ tok/s per active stream (100+ solo). Optimise aggregate throughput for the team.
Serving llama.cpp/GGUF to start → vLLM (CUDA) for multi-agent batching.
The real bill Weights aren’t the cost — KV cache for long repo context is. Size memory for context + batching.
Tier (from the article)
T1 · Solo
T2 · Agent team
T3 · Shared server
Platform
DIY · NVIDIA
Apple Silicon
T1 GPU
RTX 5070 Ti 16GB
RTX 5080 16GB
Reuse podmaster parts
On
Reuse on: podmaster already has CPU/mainboard/NVMe/case + screen, keyboard, mouse, and the wired LAN switch — only new parts are billed.
What it runs & article fit
Buy list & PhongVu sources
phongvu.vn · 2026-06-16
Component Price Source To purchase —
Filtered to the current selection above. Confirmed = live phongvu.vn page (click ↗); est = no firm SKU yet, estimate to confirm.
Prices captured 2026-06-16 from phongvu.vn (RTX 5070 Ti 35.59M, RTX 5080 44.99M, RTX 5090 32GB 119.99M, RTX PRO 6000 96GB 405.9M, Mac Studio M4 Max 64GB 82.99M, Mac Studio M3 Ultra 96GB 116.49M, Ryzen 9 7900 10.49M / 9900X 12.7M, B650 7.49M, 64GB DDR5-6000 6.99M, Corsair RM1000e 4.39M). FX 1 USD ≈ 26,300 ₫ (Investing.com, 2026-06-16). 64GB DDR4 = matched 2×32 DDR4-3200 kit 11.5M (Gigabyte Designare, listed). DDR4 is EOL-price-spiked (a single 16GB stick is ~3.79M), so 64GB DDR4 (11.5M) now costs far more than 64GB DDR5-6000 (6.99M) — reusing podmaster’s DDR4 platform saves on CPU+mainboard, not RAM. Estimated lines (est ): Ryzen 7 7700 ~7M, B650M ~4.5M, 750W PSU ~2.8M, ATX case ~2M, 2TB Gen4 NVMe ~4.5M.
Reuse baseline = podmaster : Intel i5-12400F, Gigabyte B760M DS3H DDR4, 32GB DDR4, existing NVMe, case, display/keyboard/mouse, wired LAN switch. For GPU-resident inference the CPU barely matters, so reuse is the cost-effective DIY path. Apple boxes are complete units (reuse N/A) and run MLX/llama.cpp — not vLLM/CUDA. Grounded against the article .