$5,800 Flagship Build (Budget: ~$5,800)

As of 2026-10-07. Prices are the 2026-09-28 snapshot (weekly snapshot); board numbers are sourced rows only. On our board the RTX 5090 carries 23 sourced rows out of 55 (the rest are partial), and its fastest sourced row is 446.7 tok/s (Nemotron-3.5-Lightning-30B-A3B, NVFP4, 65K). The fastest number anywhere on the board is 4,546.8 tok/s — a partial row on this card whose source never stated an engine version, so it counts in nothing. This is the build you buy when 24GB ran out of room and you want the whole 32GB tier at NVFP4 speed.

The list

What it actually runs (board rows, per-row attribution)

• Qwen3.8-27B — 435.1 tok/s (IQ4_XS, lucebox, 131K context) — the fastest sourced 27B row. A 4,546.8 tok/s NVFP4/sglang row exists in the database but is partial, so it is not a board result; the page below still shows what a finished NVFP4 stack looks like.
• LFM2-350M NVFP4A16 — 2,231.4 tok/s (partial: no engine version) — the fastest number recorded on this card; it is excluded from every aggregate, which is why the sourced ceiling below is what it is.
• Nemotron-3.5-Lightning-30B-A3B NVFP4 — 446.7 tok/s (imported, vLLM) — 30B-class MoE at real-time-plus.
• NVIDIA-Nemotron-3.5-Lightning-30B-A3B — 446.7 tok/s (NVFP4, vLLM, 65K context) — the everyday assistant class, and a sourced row.
• Kimi-Linear-48B-A3B-Instruct — 280.8 tok/s (IQ4_XS, llama.cpp, 32K) — how far the 32GB tier stretches into MoE territory.
• Kimi-Linear-48B-A3B IQ4_XS — 280.8 tok/s (imported) — a 48B-total MoE that fits 32GB because only ~3B is active.

Why this build, not two 24GB cards

• 32GB single card beats 2×24GB on bandwidth per dollar at 27B/30B sizes, with zero NVLink/P2P setup.
• NVFP4 is the native path — but read the ROI carefully. On the same model (Qwen3.8-27B, sourced rows only) the 5090's best is 435.1 tok/s against the single 3090's 164.6 tok/s: 2.6×, not an order of magnitude. The 5090's real advantage is that 27B at Q4 fits with a long context, and that Blackwell-only formats exist at all.
• Headroom: 48B-A3B-class MoE fits with room for context, which neither 24GB card nor one A6000 (48GB, but 768GB/s vs the 5090's 1.8TB/s) can match on speed.

Honest caveats

• Price volatility is the story. This build was first published when the card listed at ~$5,000; the 2026-09-28 snapshot has it at $9,000, which is why the title no longer carries a dollar figure. The weekly snapshot is the number to budget from, not this page.
• "Sourced" on this board means the row states engine and version, quantisation, context length, a specific hardware configuration, and a link to the author's original page. DiggTop does not re-run other people's measurements: 466 of the 468 rows are attributed claims from public sources, and only two rows are our own measurements (self-measured) — an RTX 4090 D at 67.4 and 70.9 tok/s. Treat every 5090 number as author-reported with a source link.
• 575W TDP: this build adds ~0.7kW of cooling load at peak and is not a quiet machine.
• If your ceiling is 32B dense, the ~$1,500 used-3090 build does 81.9 tok/s on the same 27B model for 1/4 of the money. Buy the 5090 for speed and MoE headroom, not for 32B.

Sources

• 5090 pricing: weekly price snapshot (US street average, 2026-09)
• Benchmark rows: RTX 5090 on the board — every figure above links to its row's attribution
• Speed comparison: RTX 5090 vs RTX 3090 guide