5090_ASBUILT_BENCHMARKS.md

pipelines/5090_ASBUILT_BENCHMARKS.md

THE 5090 BOX AS-BUILT BENCHMARK RECORD — Puget Order #293544

Status: ACTIVE (director-transcribed 2026-07-29 from Josh's ten photos of the Puget order page's

Benchmarks tab + five build photos of the Pictures tab; benchmarked by Puget 07/28/26 12:45 PM).

This is the QC BASELINE the machine shipped with — the arrival soak checks and every future

"is the box healthy" question compare against THIS, not against spec sheets. Archival-fidelity

source: the order page itself (Puget account, Benchmarks tab) can be saved to PDF later; the

numbers below were read from photos and the load-bearing ones are exact.

The two numbers the walk-through needed

• POWER: 155W idle / 1000W FULL LOAD (Puget-measured, stated for UPS sizing). ✅ UPS VERIFIED

(2026-07-29, arrived): CyberPower PR1500RTXL2UC — 1500VA/1500W pure sine wave, AVR, 8 outlets,

2U rack/tower, EBM-expandable. 1500W ≥ 1000W with 500W headroom (~67% worst-case load) —

properly sized; EBM batteries deferred (they add RUNTIME, not capacity; the base unit covers

the auto-shutdown window). Walk-through item 6 reduces to: seat the USB cable, install

PowerPanel, run the pull-the-plug test. (SHIPPED-CONFIRMED 2026-07-29; the box arrives

2026-07-30, a day ahead of the planned Thursday.)

• THERMAL BASELINE (under Puget's benchmark load, ambient ~25C): CPU spikes to ~90C+ on

all-core work, GPU plateaus ~80C. Puget's own note: this chart exists as the

compare-on-arrival baseline. Those CPU spikes are NORMAL for a 9950X3D under benchmark

load — nobody panic-tunes cooling on day one. The soak cadence reads against these curves.

Identity (from the build photos + labels)

• Puget Workstation Ryzen X870E R132-XL, S/N 293544-01. Windows 11 Pro, OOBE ship state.
• Chassis: Fractal Define 7 class (sound-damped front door, dust-filtered intake), LG Blu-ray

M-DISC drive, Noctua case fans throughout, AIO liquid CPU cooler (front radiator), ASUS

ProArt X870E-Creator motherboard, ASUS TUF RTX 5090, acrylic GPU support bracket.

• Rear I/O confirms: 10G + 2.5G Ethernet ONBOARD (the 10GbE assumption in the network plan

holds), Wi-Fi 7, USB4 40G + 20G + multiple 10G ports.

GPU — RTX 5090 (stock-healthy, no throttle signature)

• Nbody CUDA bandwidth: Host→Device 48,619.1 MB/s · Device→Host 55,759.1 MB/s (PCIe Gen5 x16

running at full width — matters for local-inference model loading).

• Unigine Superposition 1.0: 1080p Extreme 29,220 (218.6 avg FPS) · 1080p High 51,154

(382.6 FPS) · 4K Optimized 44,737 (334.6 FPS) · 8K Optimized 18,991 (142.0 FPS).

• OctaneBench 2025.2.1: total 1725.2.
• V-Ray 6.00.01: CPU 50,543 vsamples · CUDA (GPU) 11,242 vpaths · RTX mode 15,357 vpaths.
• KeyShot 15.0.0.174: GPU 478.82 (CPU 6.25).

CPU — Ryzen 9 9950X3D (right on the rails)

• Cinebench 2026: single-thread 549.18 · multi-core 9,757.93 (five runs 9,746–9,784, run

deviation 15.069 — tight, thermally stable).

• NeatBench 6.0.5: best combined CPU+GPU 68.35 fps (GPU-only 60.41; CPU-only peak 18.56;

per-core scaling recorded 1→32 cores on the order page).

• Lame 3.100 MP3: 60MB wav→mp3 2.8s default / 1.2s fast.

Storage — three tiers, and C: is a Gen5 surprise

• C: Kingston Fury Renegade (SFYR2S4T0) 4TB — MEASURED Gen5-class: Sequential Read (Q8T1)

14,516.427 MB/s · Sequential Write ~14,124 MB/s · Random Read (Q32T1) 209,908.7 IOPS.

A materially better boot/apps drive than the runbook's Gen4 assumption.

• D: Samsung 990 PRO 4TB — Seq Read 6,560.983 / Write 6,421.522 MB/s · Random Read (Q32T1)

144,071.5 IOPS (Gen4 normal). The work drive (repos, models, active projects).

• E: Samsung 870 QVO 8TB SATA — Seq ~563.8/535.1 MB/s (SATA normal). Bulk/archive.

Creative-suite proof points (not our workload; proves the machine end to end)

• PugetBench Photoshop 1.0.6 (PS 27.4): overall 12,737 (General 129 / Filter 126).
• PugetBench After Effects 1.1.0 (AE 26.0): overall 17,587 standard / 11,188 extended

(2D 206 · 3D 210 · Tracking 128).

• PugetBench Premiere Pro 2.0.1 (PP 26.0): overall 171,102 standard / 131,922 extended

(as printed; sub-scores LongGOP 143 · Intraframe 117 · RAW 214 · GPU Effects 238); 8K RED

processing 91.92 fps · 4K H.264 150Mbps processing 212.92 fps · ProRes 422 Proxy UHD

encode 271.45 fps.

What this changes

• UPS sizing gets its real number (1000W full load) — walk-through item 6.
• The soak/thermal reference curves exist BEFORE first boot — arrival checks compare, never guess.
• The C: Gen5 measurement upgrades the storage plan's assumptions (boot/apps I/O is not the

bottleneck anywhere).

• Benchmark day starts from a KNOWN-GOOD machine: any regression on our own runs (Cinebench,

Octane, disk) is environmental, not silicon.

Generated by harness/site/structure_site.py — the URL path is the repo path. review root