🔥 Fleet Thermal, Power & PPD Efficiency Across Hardware Eras

This engineering study cross-references 588,289 high-resolution neumon JSON telemetry records (442,962 active-folding snapshots across itztli, origami, smallpieces, scraps, and foldimate) with authoritative client.db / fah_state.json physical GPU UUID Work Unit benchmarks across all 13 hardware eras documented in Fleet Hardware Inventory & Production Over Time. Every wattage, thermal, frequency, and efficiency figure below is derived from direct silicon telemetry (NVIDIA NVML board power + AMD/Intel RAPL package counters + credited WU completion timings).

---

1. Master 13-Era Fleet Power, Heat Dissipation & PPD Efficiency Timeline

Across the 13 hardware eras from February to October 2026, fleet electrical and thermal metrics are tracked at three boundaries: Active GPU Silicon Power (sum of NVML board power across all folding GPUs), Total Silicon Power (GPU board power + CPU RAPL package power), and Estimated AC Wall Power & Thermal Output (accounting for 80+ Titanium/Platinum PSU conversion efficiency and motherboard/fan/backplane overhead; 1W = 3.41214 BTU/hr). Eras 1–4 reflect pre-neumon hardware baselines reconstructed from GPU/CPU TDP profiles and EOC logs; Eras 5–13 reflect direct neumon.log continuous telemetry.

---

2. Host CPU Thermal Evolution, Vcore/Frequency Tuning & Cooling Architecture

While GPUs execute the bulk of OpenMM molecular dynamics math, host CPUs manage PCIe DMA transfers, frame-boundary coordinate reductions, and 5%-checkpoint sanity verifications. Across the fleet, our telemetry demonstrates that targeted CPU power/heat-saving measures—such as running the Ryzen 9 9900X in 65W Eco Mode (88.0W PPT) or pinning the Kaby Lake i7-7700 to 2.40–2.80 GHz—slash CPU heat and power draw by 45% to 83% while preserving 96.7% to 100% of peak GPU PPD.

🔬 2.1 CPU Power & Frequency Tuning vs GPU PPD Retention

• smallpieces v3 — AMD Ryzen 9 9900X in 65W Eco Mode (88.0W PPT vs 162W Stock PPT):
  • To keep CPU heat dissipation compatible with the low-profile 53mm Thermalright AXP90-X53 Full Copper cooler inside the vertically oriented GPNE wind-tunnel frame, the 12-core / 24-thread Ryzen 9 9900X (120W TDP / 162W stock PPT ceiling) is configured in UEFI 65W Eco Mode (88.0W PPT limit, a -45.7% reduction in peak package power ceiling).
  • Under full 5-GPU Blackwell load (45.33M PPD), neumon records the Ryzen 9 9900X at 88.0W RAPL, 1.055V Vcore, 65.8°C mean (66.8°C p95), and only 22.2% thread utilization across 24 threads (down from 93.2% on 4-core smallpieces v2). Despite the 88.0W PPT cap, smallpieces v3's MSI RTX 5070 achieves the fastest Project 18257 frame time in the fleet (36.58s TPF / 13.01M PPD, slightly ahead of origami v2's uncapped 32-core EPYC at 37.08s / 12.45M PPD)—proving that 65W Eco Mode sacrifices 0.0% of GPU folding throughput while saving up to 74W (252 BTU/hr) of CPU heat.
• smallpieces v2 & scraps v2 — Octominer B250 Intel Core i7-7700 MSR 0x1a0 Unlock & 2.40–2.80 GHz Sweet Zone:
  • The 800 MHz BIOS Trap: The Octominer B250 mining motherboard boots with IA32_MISC_ENABLE (MSR 0x1a0) bit 38 set (0x4000850089, Turbo Boost disabled), trapping intel_pstate at 800 MHz (8x multiplier, 0.672V, 4.3W) 93% of the time and inflicting a +4s to +5s GPU TPF penalty (-13.2% frame speed).
  • The BIOS "Turbo Performance" Voltage Spike & Cooler Investigation (86–89°C): Enabling "Boot performance mode = Turbo Performance" in the AMI BIOS forces the 36x (3.60 GHz) state at an aggressive 1.140V Vcore (25.3W RAPL), driving CPU temperatures to 83–89°C and screaming the four Octominer Delta server fans at 85% PWM (4,700 RPM). On August 8, 2026, swapping the 2-heatpipe Noctua NH-L9i (37mm) for a 4-heatpipe Thermalright AXP90-X36 (36mm) + Arctic P8 PWM PST scavenger fan proved that both coolers perform identically within <0.5°C (51.2–52.1°C mean) once the BIOS Vcore spike is eliminated.
  • The 2.40 GHz (10.4W) & 2.80 GHz (14.0W) Sweet Spots: Using nvpowerctl v1.3.2 to clear MSR 0x1a0 bit 38 at boot and pin cpu_max_freq_khz to 2400000 (2.40 GHz, 0.864V, 10.4W, 53.5–59.0°C) or 2800000 (2.80 GHz, 0.944V, 14.0W, 64.0°C) cuts CPU power by -44.7% to -58.9% vs 3.60 GHz (and by -82.8% vs scraps v1's Haswell i7-4790 at 60.4W) while retaining 96.7% (at 2.40 GHz) to 98.0% (at 2.80 GHz) of the maximum 3.60 GHz GPU frame speed and keeping the Octominer Delta fan wall at its whisper-quiet 22% PWM (1,805 RPM) floor:

🌀 2.2 Complete Fleet CPU & Chassis Cooling Solutions Inventory

---

3. PCIe Generation, Lane Width & CPU Feeder Scaling: smallpieces (v2 → v3), origami (v1 → v2) & scraps (v1 → v2)

Across the fleet's history, three major platform rebuilds transplanted the exact same physical GPU silicon UUIDs onto faster PCIe buses and higher-core-count CPUs. Together, they provide definitive proof of how PCIe link bandwidth (Gen 2.0 x1 vs Gen 3.0 x1 vs Gen 3.0 x16 vs Gen 4.0 x16 / Gen 5.0 x4) and CPU feeder headroom govern both GPU power cap utilization and Folding@Home PPD.

Hardware Terminology Note on Contiguous Mining Motherboards vs PCIe Cable Risers: Ribbon or cable-connected PCIe risers never existed in any fleet build prior to smallpieces v3. Both the AX-B85M-ETH (smallpieces v1 and scraps v1) and the Octominer X12 Ultra (smallpieces v2 and scraps v2) seat GPUs directly into rigid, board-mounted physical x16 slots—on the AX-B85M-ETH via a super-extended contiguous motherboard wired as PCIe Gen 2.0 x1, and on the Octominer via an ATX-style B250 motherboard whose edge connector mates directly into a rigid powered PCIe extension backplane (Slot 1 wired as PCIe Gen 3.0 x16; Slots 2–12 wired as PCIe Gen 3.0 x1). Only smallpieces v3 uses cable risers (four shielded ADT-Link K43SM (2x 40cm + 2x 50cm) M.2-to-PCIe Gen 5.0 x4 risers off the passive ADT-Link R34A bifurcation card in PCIEX16_1, while the 5th GPU in PCIEX16_2 is seated directly in the motherboard slot without a riser).

🚀 3.1 The Bus-Starvation Wattage Paradox on smallpieces v2 → v3 (+102.9% PPD, +47.3% PPD/W)

• Why PCIe x1 Backplane + 4-Core CPU Starvation Masquerades as "Low Power Draw": On smallpieces v2, four RTX 5060 Ti cards sat in Octominer PCIe Gen 3.0 x1 powered backplane slots (~985 MB/s link bandwidth) while sharing a 4-core / 8-thread i7-7700 CPU running at 93.2% utilization across 6 concurrent GPUs. At every OpenMM frame boundary and PME reduction step, the GPUs stalled waiting for PCIe DMA transfers and CPU thread scheduling. Because GPU shaders drop to low-power idle states during host/bus stalls, the 150W-capped RTX 5060 Ti cards averaged only 96.6–103.1W (64–69% of cap) and the 175W-capped RTX 5070 (in Slot 1 Gen 3 x16, but CPU-starved) averaged 139.0W (79% of cap).
• Unleashing Full Silicon Utilization on smallpieces v3: Moving those 5 GPUs onto an ADT-Link R34A x4/x4/x4/x4 bifurcation board (4x K43SM (2x 40cm + 2x 50cm) risers at PCIe Gen 5.0 x4 = ~15.75 GB/s per card, 16x the bandwidth of Gen 3 x1) plus PCIEX16_2 direct on the motherboard at PCIe Gen 4.0 x4 (~7.88 GB/s) and a 12-core / 24-thread Ryzen 9 9900X (88.0W Eco Mode) eliminated all DMA and CPU feeder stalls. Each RTX 5060 Ti immediately rose to 135.8–147.1W (91–98% of cap) and the RTX 5070 rose to 174.2W (99.5% of cap).
• Super-Linear Efficiency Gain (+38.3% to +60.7% Higher PPD/W): Because a stalled GPU still burns ~35–45W of fixed board/VRAM/PLL baseline power while producing zero math during DMA waits, eliminating bus stalls increased output (+102.9% PPD, 22.34M → 45.33M PPD) 2.5x faster than power draw (+40.3% W, 507.6W → 712.4W mean). The 5-GPU smallpieces v3 array produces +64.2% more PPD than the 6-GPU smallpieces v2 array (which also had a ZOTAC RTX 4070 Blower installed).

⚡ 3.2 origami v1 (i9-7940X PCIe Gen 3.0 x16) → origami v2 (EPYC 7543 PCIe Gen 4.0 x16): +8.0% to +13.8% PPD & Cooler Temps

Even when GPUs are already seated in full x16 motherboard slots, upgrading from a 14-core Skylake-X i9-7940X (PCIe Gen 3.0 x16 = 15.75 GB/s, 19.25 MB L3 cache, mesh interconnect) on origami v1 to a 32-core Milan AMD EPYC 7543 (PCIe Gen 4.0 x16 = 31.5 GB/s, 256 MB L3 cache) on origami v2 produced a measurable, across-the-board speedup on the exact same physical NVIDIA RTX PRO 4000 Blackwell #1 (6f99ccc4, 145W cap) and ZOTAC GeForce RTX 5070 Solid OC (dd40d876, 180W cap) across 100,870+ neumon telemetry records and every matched Work Unit project:

• Higher Cap Utilization at Lower GPU Temperatures: On origami v2, RTX PRO 4000 Blackwell #1 (6f99ccc4) power draw rose from 139.5W (96.2% of cap, n=42,076) to 143.6W (99.0% of cap, +4.1W, n=58,794) while its mean core temperature dropped by -4.4°C (78.8°C → 74.4°C) thanks to the Phanteks Enthoo Pro 2's 13x Super Flower MEGACOOL fan array and dedicated vertical inter-GPU fan. Simultaneously, Zotac RTX 5070 (dd40d876) power draw rose from 175.4W (97.4% of cap, n=42,145) to 178.7W (99.3% of cap, +3.3W, n=59,259) while running -2.5°C cooler (61.0°C → 58.5°C).
• -5.2% to -12.4% Faster TPF & +8.0% to +13.8% Higher PPD Across Every Matched Project: Doubling slot bandwidth from PCIe Gen 3.0 x16 to PCIe Gen 4.0 x16 and expanding L3 cache 13.3x (19.25 MB → 256 MB) shaved 2.0s to 9.0s off every frame across all four projects that ran on both versions of origami.
• scraps v1 (AX-B85M-ETH PCIe Gen 2.0 x1) → scraps v2 (Octominer PCIe Gen 3.0 x1 Backplane): Doubling physical x1 slot bandwidth from 500 MB/s (Gen 2.0 x1) to 985 MB/s (Gen 3.0 x1) on the three matched RTX 3060 Ti GPUs reduced Project 15506 TPF by -6.8% (59.47s → 55.4s) and boosted per-card PPD by +11.0% to +15.8% (2.83M → 3.14M–3.28M PPD).

---

4. GPU Architecture & Silicon Efficiency Matrix (Ada vs Blackwell vs Ampere vs Turing)

Correlating per-GPU NVML power draw and core temperatures from neumon with credited Work Unit PPD across all 18 physical GPUs that have operated in the fleet reveals a 6.0x spread in silicon efficiency between the flagship Ada Lovelace RTX 4090 (127.2k PPD/W) and the retired Turing RTX 2060 (21.0k PPD/W).

• Tier 1 — Flagship Ada Lovelace & Workstation Blackwell (89.8k – 127.2k PPD/W):
  • NVIDIA GeForce RTX 4090 (itztli, 315W cap, 261.0W mean active draw @ 58.1°C): 33.20M PPD at 127.2k PPD/W (peaking above 36M PPD / 135k PPD/W on large beta WUs). Operating the AD102 die at ~58% of its 450W stock TGP places it deep in its voltage-efficiency sweet spot.
  • 2x NVIDIA RTX PRO 4000 Blackwell (origami v2, 145W cap, 141.7–143.6W mean draw @ 67.4–74.4°C): 12.86M–12.88M PPD each at 89.8k–90.8k PPD/W in single-slot workstation blower form factors.
• Tier 2 — Efficiency-Capped Consumer Blackwell & Ada (54.0k – 74.7k PPD/W):
  • MSI RTX 5070 VENTUS 2X OC (smallpieces v3, 174.2W @ 63.7°C) & ZOTAC RTX 5070 Solid OC (origami v2, 178.7W @ 58.5°C): 12.31M–13.01M PPD at 69.5k–74.7k PPD/W.
  • ASUS RTX 4070 Ti (foldimate, 195W shift cap, 184.7W mean draw @ 71.2°C): 11.50M PPD at 62.3k PPD/W.
  • 4x RTX 5060 Ti on smallpieces v3 (150W cap, 142.2W mean draw @ 65.8°C): 7.88M–8.27M PPD each at 54.0k–60.9k PPD/W.
• Tier 3 — Ampere Workhorses on scraps v2 (26.1k – 31.8k PPD/W):
  • ZOTAC RTX 3070 (145W cap, 141.6W mean draw @ 60.9°C in Octominer Slot 1 Gen 3 x16): 4.50M PPD peak / 3.59M PPD multi-project avg at 25.4k–31.8k PPD/W.
  • 3x RTX 3060 Ti (125W cap, 119.8W mean draw @ 57.1°C in Octominer Gen 3 x1 backplane Slots 4, 7, 10): 3.10M–3.28M PPD each at 26.1k–27.4k PPD/W (up +11.5% in efficiency over scraps v1's Gen 2 x1 motherboard).
• Tier 4 — Retired Turing Cards (21.0k – 22.2k PPD/W — De-Allocated Aug 29, 2026):
  • Gigabyte RTX 2060 SUPER (117.8W @ 60.4°C, 2.61M PPD, 22.2k PPD/W) and ASUS RTX 2060 (117.2W @ 66.0°C, 2.46M PPD, 21.0k PPD/W) burned 235.0W of continuous GPU power (~285W at the wall / 973 BTU/hr) for just 5.07M PPD combined—less than 40% of a single 142W RTX PRO 4000 Blackwell. Retiring them in Era 11 raised fleet-wide GPU efficiency by +24.3% overnight.

---

5. Controlled RAM Speed Evaluation (origami v2 A/B Test, Planned smallpieces v3 Test & Host RAM Capacity Cliffs)

Evaluating whether system DRAM clock speed affects GPU Folding@Home PPD requires strict apples-to-apples hardware isolation—holding the CPU, motherboard, PCIe lane topology, and physical GPU silicon UUIDs constant while varying only memory frequency. Cross-host comparisons (such as comparing scraps v1 vs v2 or origami v2 vs smallpieces v3) are confounded by simultaneous changes in PCIe generation, CPU IPC, and GPU boost curves and therefore cannot isolate RAM speed.

🧪 5.1 Controlled Apples-to-Apples A/B Test: origami v2 8-Channel DDR4-2400 vs DDR4-3200 (+33.3% RAM Speed)

When origami v2 (AMD EPYC 7543 + ASRock Rack ROMED8-2T + 2x RTX PRO 4000 Blackwell + 1x Zotac RTX 5070) first launched on August 16, 2026, its 256GB memory kit contained seven Samsung M393A4K40BB1-CRC (DDR4-2400) RDIMMs and one DDR4-3200 module, forcing all 8 memory channels to train at 2400 MT/s (153.6 GB/s peak theoretical bandwidth) through August 20. Between August 21 and September 5, the mismatched kit was RMA-replaced with eight matched 32GB Micron 36ASF4G72PZ-3G2R1 DDR4-3200 ECC RDIMMs running in 8-channel mode at 3200 MT/s (204.8 GB/s — exactly +33.3% higher memory clock and bandwidth) on the exact same CPU, motherboard, PCIe 4.0 x16 slots, and GPUs:

🔬 5.2 Scope of the 8-Channel EPYC Result & Planned 2-Channel smallpieces v3 A/B Experiment (DDR5-6000 → DDR5-4800)

• Why origami v2 Only Bounds High-Bandwidth 8-Channel Platforms: Because origami v2 uses an 8-channel server memory bus, even its slower DDR4-2400 configuration provides 153.6 GB/s of theoretical bandwidth (already higher than a dual-channel DDR5-6000 desktop bus at 96.0 GB/s) alongside a massive 256 MB L3 cache. Thus, origami v2 proves that raising system memory bandwidth above 153.6 GB/s yields zero F@H frame speedup while costing +15.7W of memory-controller/Infinity-Fabric power—but it does not by itself rule out RAM clock sensitivity on a 2-channel desktop platform feeding 5 GPUs.
• Planned Follow-Up Controlled Experiment on smallpieces v3 (DDR5-6000 vs DDR5-4800): To obtain an equally rigorous apples-to-apples test on a 2-channel desktop platform, smallpieces v3 (Ryzen 9 9900X + 5 Blackwell GPUs) can be stepped down in UEFI from DDR5-6000 (96.0 GB/s) to DDR5-4800 (76.8 GB/s, -20% bandwidth) with all other hardware and power caps held constant, comparing matched-project TPF, PPD, and Ryzen 9 9900X SoC/UMC power draw.
• Host RAM Capacity & Allocation Cliffs (Distinct from RAM Clock Speed):
  1. The 8GB RAM Swap Cliff (AX-B85M-ETH on smallpieces v1 & scraps v1): When 4–5 GPUs folded simultaneously on the legacy 8GB DDR3 motherboard, host memory pressure triggered kswapd0 and zswap disk thrashing ("iowait hell"), stalling GPU feeders and cutting PPD by 25–50%.
  2. OpenMM Hidden CPU Sanity-Check Thread Inflation (OPENMM_CPU_THREADS=4): Every 5% of a Work Unit, FahCore_27 initializes a hidden OpenMM CPU platform context to sanity-check GPU forces. By default, OpenMM spawns 4 × N_CPU_threads + 8 threads and heap buffers per GPU—consuming 25.95 GiB RSS across 5 GPUs on 24-thread smallpieces v3 (81% of 32GB RAM). Setting OPENMM_CPU_THREADS=4 in /opt/docker-configs/fah-compose.yaml slashed 5-GPU host RSS to ~14.0 GiB (-12 GiB saved) with 0.0% TPF or PPD impact.

---

6. Datacloset Thermodynamics, HVAC Zone Partitioning (1-Ton Mini-Split vs Attic Chimney vs 3-Ton Carrier Heat Pump), 240V Cutover & Audimute Acoustic/Thermal Isolation

Silicon efficiency only tells half the thermal story; the other half is how 10,220 BTU/hr (~2,995W AC wall) of concurrent fleet heat is physically evacuated without overwhelming room ambient temperatures or household acoustics. The fleet is partitioned across three distinct physical cooling zones:

1. The Datacloset Room Air Zone (13' H × 7' W × 4' D = 364 cu ft / 10.31 m³): Cooled by a dedicated 1-ton (12,000 BTU/hr / 3,517W thermal) mini-split AC (live May 15, 2026) assisted by a bottom door inlet louver/vent (centered in the 7 ft wall) and a ceiling Panasonic WhisperQuiet 110 CFM bathroom exhaust fan (providing 18.1 air changes per hour [ACH]). In addition to its assigned folding servers, the datacloset houses an always-on ~125W (427 BTU/hr) infrastructure rack: 1x NAS, 2x ISP modems, 1x Netgate 4100 firewall, 1x Intel NUC (m), 1x 24-port 1G switch, and 2x mesh Wi-Fi nodes (1 trusted, 1 untrusted).
2. The Datacloset Attic Chimney Exhaust Zone: A vertical chimney duct inside the datacloset captures hot exhaust directly off the Octominer server chassis and vents it straight into the attic, where a thermostatically controlled attic fan evacuates stored heat outdoors through gable vents—preventing ~85% of the chimney-ducted host's heat (~1,711–2,360 BTU/hr) from ever entering the datacloset room air or loading the 1-ton mini-split.
3. The Main House Zone (itztli & foldimate): Cooled by the main house HVAC—originally traditional forced-air HVAC, and since May 15, 2026 a whole-house 3-ton (36,000 BTU/hr) Carrier high-efficiency heat pump that went online on the same day as the datacloset mini-split.

🏗️ Chronological Facility, Electrical (240V) & Acoustic (Audimute) Engineering Milestones

• Pre–May 15, 2026 (115V Garage + Exhaust-Only Datacloset): Prior to May 15, origami v1 and the network/NAS gear (~2,798 BTU/hr total) operated in the datacloset cooled solely by make-up house air through the door inlet louver and the ceiling 110 CFM WhisperQuiet exhaust fan, while smallpieces v1 ran in the garage on 115V power (~2,610 BTU/hr).
• May 15, 2026 (1-Ton Datacloset Mini-Split + Attic Chimney + 3-Ton Main-House Carrier Heat Pump Online): Commissioning the 1-ton (12,000 BTU/hr) mini-split allowed smallpieces v1 to move from the garage into the datacloset on the attic chimney duct. Because the chimney + attic gable fan evacuated ~2,220 BTU/hr directly into the attic, the 1-ton mini-split only had to handle origami v1 (~2,371 BTU/hr) + the network/NAS stack (427 BTU/hr) + ~15% chimney conductive bleed (~332 BTU/hr) = 3,130 BTU/hr (just 26.1% of the mini-split's 1-ton capacity).
• June 27, 2026 (240V Power Energized in the Datacloset): Three days prior to the June 30 smallpieces v2 Octominer X12 Ultra cutover, dedicated 240V circuits went live in the datacloset. Running origami's Corsair AX1600i (80+ Titanium) and the Octominer's 4x HPE 750W Platinum Plus server PSUs on 240V instead of 115V improved AC-to-DC PSU conversion efficiency by +2–3% (shedding ~50–65W / ~190 BTU/hr of waste PSU heat) and cut branch-circuit current in half.
• July 25, 2026 (scraps v1 Added Un-Ducted in Datacloset Room Air — +2,330 BTU/hr Room Spike): scraps has only ever operated in the datacloset. Because smallpieces v2 occupied the attic chimney, bringing scraps v1 (5 GPUs + Haswell i7-4790 in the open GPNE frame, ~725W wall / 2,474 BTU/hr) online discharged its heat directly into the room air alongside origami v1—raising datacloset room load from 3,150 to 5,480 BTU/hr (45.7% of 1-ton capacity). In neumon telemetry (Jul 25–30), origami v1's mean GPU temperature immediately rose by +1.6°C (66.3°C → 67.9°C) and its 240mm AIO CPU temperature rose by +3.7°C (53.9°C → 57.6°C).
• July 31, 2026 (Audimute Acoustic Lining & Perimeter Door Seal Installed): To isolate server fan noise from the living space, the 13' H × 7' W × 4' D datacloset was lined with three commercial Audimute acoustic treatments:
  1. AcoustiFelt™ Fabric Acoustic Shapes (SKU: FH-400-400-C, Aluminum Felt Planks, 1.5" thickness, 0.95 NRC, Qty 6 packs): Absorbs 95% of incident mid/high-frequency acoustic energy (500–4,000 Hz), eliminating parallel-wall flutter echo across the narrow 4 ft room depth.
  2. Peacemaker® Sound Barrier — 3 mm (SKU: APM64000851, 2' × 18' = 36 sq ft high-density recycled rubber barrier roll): Adds mass-loaded transmission loss against low-frequency chassis resonance.
  3. Acoustic Door Seal Kit (SKU: ASLK0000780, Standard Interior Door Perimeter Seal): Seals acoustic leaks around the centered interior door perimeter while channeling make-up airflow strictly through the lower door inlet louver.
  Measured Thermal Side-Effect of Acoustic Insulation (Jul 31 – Aug 15): Because 1.5" of dense AcoustiFelt + 3mm rubber barrier adds ~R-5.5 of thermal insulation over the drywall and seals door perimeter air gaps, conductive heat transfer through the closet walls into the house dropped to near zero. Retaining 100% of the heat inside the closet envelope for the mini-split and 110 CFM ceiling fan nudged origami v1's GPU mean up by +0.5°C (67.9°C → 68.4°C) and CPU mean up by +1.8°C (57.6°C → 59.4°C).
• October 26–29, 2026 (The Great Chimney Swap: scraps v2 Now Chimney-Vented; origami v2 + smallpieces v3 Cooled by the 1-Ton Mini-Split): Transplanting scraps v2 (1x RTX 3070 + 3x RTX 3060 Ti, ~590W wall / 2,013 BTU/hr) into the enclosed Octominer X12 Ultra placed scraps v2 on the attic chimney duct (evacuating ~1,711 BTU/hr to the attic and dropping scraps's CPU mean to 53.5°C). The 1-ton mini-split now cools origami v2 (2,201 BTU/hr) + smallpieces v3 (3,207 BTU/hr in the vertical GPNE wind-tunnel) + Network/NAS rack (427 BTU/hr) + 15% chimney bleed (302 BTU/hr) = 6,135 BTU/hr (51.1% of the 1-ton mini-split's 12,000 BTU/hr capacity, leaving 5,865 BTU/hr of thermal headroom).
• November 13, 2026 (smallpieces v3.1 Acoustic & PWM Upgrade: Dual Alphacool ES Guardian Hubs + 1" Audimute PSU Hush Box): Replaced the GPNE frame's manual dial fan hubs with two motherboard-controlled Alphacool ES Guardian 4-Port Powerfan Hubs (CHA_FAN2 Intake & CHA_FAN1 Exhaust on Nuvoton nct6799, driving 8x M12038 120x38mm 1.50A fans under gpufanpin v1.4.0 positional PWM curves: ~850 / 1,518 RPM off-shift idle, ~2,195 / 2,292 RPM 5-GPU load, 100% open-collector failsafe on PWM loss / S5 soft-off), and fabricated a custom 1" Audimute Eco-C-Tex acoustic hush box over the LX2400W 2400W server PSU exhaust bay to tame its high-frequency internal PSU fan whine.

---

7. Utility Time-of-Use (TOU) Duty-Cycle Shaping: Daily Energy (kWh/day), Live Baseline & Thermal Load

While instantaneous efficiency is governed by silicon power caps and PCIe bus architecture, 24-hour cumulative heat and energy consumption are governed by automated systemd shift timers (fah-start / fah-stop). During the EV2-A peak/part-peak window (15:00–00:00, 9 hours/day), Folding@Home is paused across all fleet hosts (0 W of F@H folding load), and the two high-wattage desktop rigs (itztli, foldimate) fold strictly during the coolest overnight off-peak hours (00:00–06:15, 6.25 hours/day). However, the three dedicated datacloset servers (origami, smallpieces, and scraps) remain live and online 24/7, drawing continuous baseline idle power during 15:00–00:00 while origami and smallpieces also serve intermittent, VRAM-resident multi-GPU compute workloads whose CPU and GPU power draw oscillates during active bursts:

• Peak-Window Thermal & Energy Relief (-43.9% Daily kWh & Heat vs 24/7, Shedding 83.2% of Server Load During 15:00–00:00): Pausing F@H on the three dedicated servers (origami, smallpieces, scraps) during the warmest afternoon/evening utility window (15:00–00:00) eliminates 100% of F@H folding watts and drops 3-server AC wall draw from 2,179 W (7,435 BTU/hr) to a 229.3 W quiescent silicon baseline (~350 W AC wall, ~365 W time-averaged with intermittent multi-GPU compute bursts, or 1,194–1,245 BTU/hr). Combined with running the two high-wattage desktop rigs (itztli, foldimate) strictly during the coolest overnight off-peak hours (00:00–06:15), the Era 13 fleet consumes 40.3 kWh/day (137.5 kBTU/day, including 3.3 kWh/day during the 9-hour server pause window) instead of 71.9 kWh/day (245.3 kBTU/day)—a -43.9% reduction in total daily heat and electrical energy.
• Why 60-Second neumon Telemetry Under-Samples Short Compute Oscillations: Even with 54.3 GiB of VRAM allocated on origami v2 and 49.3 GiB allocated on smallpieces v3, Blackwell GPUs drop smoothly into hardware P8 idle (2.5–13.7 W per card, 25.0–29.1 W per host) whenever compute kernels are not actively executing. Sustained multi-minute compute bursts on origami v2 are readily captured by neumon's 60-second polling interval (n=478 samples ramping to 291.9 W median / 348.7 W p90 silicon), whereas shorter sub-minute interactive bursts on smallpieces frequently complete between 60-second sample ticks (capturing n=12 mid-burst snapshots at 136.7 W median / 149.1 W max silicon on v2).
• Era 13 EV2-A Surpasses Era 9 TOU-C Output on -13.1% Less Daily Energy: Even though the EV2-A server folding window is 4 hours shorter per day than the old TOU-C schedule (15h vs 19h) and carries 4 extra hours of live server baseline draw, the Era 13 fleet produces +8.6% more net daily points (74.8M vs 68.9M PPD/day) while consuming -6.1 kWh/day (-13.1%, 40.3 vs 46.4 kWh/day) less electricity and generating -20.8 kBTU/day less heat—a +25.7% improvement in net daily energy efficiency (1.48M → 1.86M PPD/kWh).

---