Leaderboards · Mods
Mods, with regression checks.
Repad, repaste, undervolt, fan-curve A/B. Each entry pairs a thermal claim with a noise figure (or an explicit n/a-with-reason), a perf claim with a power figure, and a quality-regression check — same-workload pass-counts both sides. Single-axis claims invite gaming.
Scoring methodology: v1
Stock-configuration entries are on a separate board: /leaderboards/hardware.
thermal pad swap (front + back) + repaste
RTX 3090 FE · 2026-04-26 → 2026-05-01Seller PC (i9-12900KF / 64 GB DDR4 / B760M-A)
Significant memory-junction headroom restored; +29% sustained AI throughput on workloads that fit fully in 24 GB VRAM; no quality regression.
3DMark thermals (Side A → Side B)
Memory junction peak dropped 10 to 16 °C across the three workloads.
| Test | Side A junction | Side B junction | Δ junction | Δ hotspot | Side B power max |
|---|---|---|---|---|---|
| Time Spy | 100 °C | 84.0 °C | -16 °C | -14.3 °C | 350 W |
| Speed Way | 98 °C | 88.0 °C | -10 °C | -7.8 °C | 350 W |
| Steel Nomad | 96 °C | 86.0 °C | -10 °C | -8.2 °C | 349 W |
Sustained AI throughput
qwen2.5-coder:32b · 8 cases × 20 reps
Side A total elapsed
4053 s
Side B total elapsed
2885 s
Throughput delta
+28.8%
Quality, power, noise
Quality regression
no regression
Power held flat?
yes
Both sides pinned at 350W cap across all three 3DMark workloads.
Noise (dBA)
n/a
n/a — no calibrated dBA meter on this run; fan curve unchanged between sides.
Raw data
-
Side B headline report ·
reports/benchmarks/seller-pc-3090-fe-fresh-pads-2026-04-28/SIDE_B_FINAL_REPORT.mdview ↗ ↗ -
Side B qwen32b run JSON ·
reports/benchmarks/seller-pc-3090-fe-fresh-pads-2026-04-28/qwen32b-rep20-run1.jsonview ↗ ↗ -
Side A qwen32b run JSON ·
reports/benchmarks/seller-pc-3090-alone-large-2026-04-27/qwen32b-rep20.jsonview ↗ ↗
Methodology: /leaderboards/methodology. JSON API: /leaderboards/mods.json.