← Back to home

smallpieces Hardware, History, and Roadmap

Page Overview: Complete operational chronicle, physical slot mapping, fleet Docker Compose architecture, and upgrade roadmap for server smallpieces.

---

📜 1. Historical Journey (Feb 2026 – Present)

1.1 Executive Overview & Architectural Evolution

Server smallpieces originated as a budget multi-GPU prebuilt mining rig frame designed to process Folding@Home Work Units across cost-effective mid-tier graphics cards. Over time, it evolved from a memory-constrained Haswell system (smallpieces v1), through a 12-slot Kaby Lake server chassis (smallpieces v2, subsequently transferred to scraps v2), to its current unstarved 12-core Zen 5 + PCIe Gen 5.0/4.0 x4 5-GPU All-Blackwell architecture (smallpieces v3, live October 29, 2026).

---

1.2 Phase 1: smallpieces v1 Era (Feb 2026 – June 2026)

• Chassis & Motherboard: GPNE ATX Full Tower Steel Mining Rig Frame Case with integrated AX-B85M-ETH motherboard.
• CPU & RAM Upgrades:
  • Stock prebuilt shipped with a low-tier Celeron CPU and 4GB RAM SODIMM.
  • User Upgrade: Upgraded to an Intel Core i7-4790 (Haswell 4-Core / 8-Thread, LGA 1150) paired with Noctua NH-L9x65 cooler and an 8GB DDR3-1600 SODIMM module (the absolute single-slot hard limit of the AX-B85M-ETH motherboard).
• Power Supply & Electrical Circuit: 2400W Server Mining Power Supply operating on a standard 120V / 15A household circuit. This necessitated strict software power limits across all installed GPUs to maintain safe electrical headroom and prevent breaker trips.
• Primary Storage: 30GB mSATA SSD.

Memory Constraints, "iowait Hell", and the zswap Hack

• The Memory Crisis: Concurrent F@H "whale" jobs (>1GB RES RAM each) frequently overwhelmed the 8GB RAM limit, causing severe swapping and disk I/O thrashing ("iowait hell").
• Storage & Power Hack (March 19, 2026):
  • Increased system throughput via "Plan /dev/sdb", or, adding a secondary 2.5" Crucial M4 64GB SATA SSD and moving the folding operations I/O to this storage by migrating /opt to it.
  • Power Adapter Hack: Bypassed the undocumented JSATA_PWR motherboard header by powering the SSD via an external USB 3.0 to SATA power adapter.
• The zswap Compression Hack:
  • Configured Linux zswap in-RAM compressed swap allocation.
  • Compressed memory pages prior to swapping to disk, significantly dampening I/O latency and stabilizing 24/7 execution during heavy multi-job peaks.

Early GPU Telemetry & The AMD RX 9060 XT Test

• Early Baseline: 2x NVIDIA GeForce GTX 1070 (500fc1e1, bb7dfb34).
• The AMD Test ("Middle Ages"):
  • Added 1x AMD Radeon RX 9060 XT (38366633).
  • Purpose: Deployed specifically as an experimental testbed to validate and benchmark telemetry logging routines inside neumon and fahnalysis against AMD ROCm / OpenCL GPU architectures.
  • Outcome: Yielded poor PPD throughput compared to NVIDIA cards and was subsequently decommissioned.
• Pre-Migration Lineup: 1x RTX 4070 (12GB), 1x RTX 2060 SUPER, 1x RTX 2060, and 3x RTX 3060 Ti cards.

---

1.3 Phase 2: smallpieces v2 Migration & Octominer Discovery (June 27 – June 30, 2026)

• The Transition Rationale: The original plan was simply to relocate the v1 rig to a newly installed 240V / 30A circuit to lift power cap restrictions. However, the Haswell i7-4790 CPU bottlenecks and 8GB RAM single-slot limit were too restrictive for long-term fleet scaling. This prompted hardware research that led directly to the discovery and purchase of a surplused Octominer X12 Ultra chassis.
• The New Platform: Rebuilt inside the Octominer X12 Ultra 12-slot server chassis connected to the dedicated 240V / 30A circuit.

Key Architectural Correctives

1. Cooler Selection (Noctua NH-L9i): The planned Thermalright AXP90-X53 Full Copper cooler exceeded the lid clearance of the OX12U chassis. The system ran temporarily on a re-pasted stock cooler before installation of the low-profile Noctua NH-L9i (37mm height), which with the finstack parallel to the DRAM to prevent a heatwall, kept the CPU at 55–70°C with 4mm clearance between the fan and PSU cage.
2. RAM Bottleneck Eliminated: Upgraded to 32GB DDR3L (2x 16GB SODIMM), completely removing the legacy 8GB Haswell swap constraint.
3. Quad Power Supply Array: Powered by 4x (quad) Octominer HPE 750W Platinum Plus PSUs running on 240V input for optimal power distribution across 12 slots.

---

1.4 Phase 3: Modern Fleet Standard (July 2026 – October 25, 2026)

• Kernel & Driver Upgrade: Running Linux Kernel 7.0.0-28-generic (Ubuntu LTS 24.04.4 HWE track) with NVIDIA Driver 610.43.02.
• 6-GPU Array & RTX 5070 Installation (July 18, 2026): Installed a net-new MSI GeForce RTX 5070 VENTUS 2X OC (12GB GDDR7) in physical Slot 1 (CPU-adjacent PCIe Gen3 x16 slot), relocated the ZOTAC RTX 4070 to physical Slot 6 (x1 riser), and retained all 4x RTX 5060 Ti cards across Slots 4, 8, 10, and 12.
• Cooler Swap to Thermalright AXP90-X36 (August 2026): In August 2026, the Thermalright AXP90-X36 (36mm height, 4 heatpipes, copper heatpipe matrix) was swapped in to replace the Noctua NH-L9i in an effort to remedy a perceived sudden, inexplicable 10–20°C rise in CPU temperatures across all load levels. However, the cooler swap alone did not lower temperatures. Root cause analysis ultimately identified the motherboard BIOS setting "Boot performance mode" as the culprit: in search of core/thread performance, it had been changed to "Turbo Performance" instead of the default "Max Non-Turbo Performance", which forced elevated Vcore voltages that drove higher operating temperatures. After reverting the BIOS setting, empirical testing confirmed that the Thermalright AXP90-X36 delivers identical temperatures for the same loads as the Noctua NH-L9i. Notably, its dual heatpipe count (4 vs. 2), greater metal mass, and copper construction did not yield a measurable thermal delta over the single-pair NH-L9i design.

---

🎛️ 2. Physical to Logical GPU Ordering Map (smallpieces v3 & Historical v2)

2.1 GPU, Riser & Storage Slot Configuration (5-GPU All-Blackwell Array)

With the GPNE chassis standing vertically on its end, the 5 Blackwell GPUs are mounted with approximately 2 PCI slot widths of spacing between each card from bottom to top: Bottom (1st): PNY RTX 5060 Ti 16GB VERTO (R34A Slot 4, GPU 3) ➔ 2nd from bottom: MSI RTX 5060 Ti 8G SHADOW #1 (R34A Slot 3, GPU 2) ➔ Middle (3rd): MSI RTX 5060 Ti 8G VENTUS (R34A Slot 2, GPU 1) ➔ 4th from bottom: MSI RTX 5070 12GB VENTUS 2X OC (R34A Slot 1, GPU 0) ➔ Top (5th): MSI RTX 5060 Ti 8G SHADOW #2 (GPU 4, plugged straight into motherboard slot PCIEX16_2 without a riser). Primary CPU slot PCIEX16_1 is bifurcated x4/x4/x4/x4 in UEFI across the passive ADT-Link R34A quad M.2 adapter and 4x ADT-Link K43SM M.2-to-PCIe x16 risers (2x 40cm + 2x 50cm, all four negotiating at the motherboard's native PCIe Gen 5.0 x4, 32 GT/s), while the non-riser card in PCIEX16_2 negotiates at native PCIe Gen 4.0 x4 (16 GT/s). There are no x16-to-x16 nor x4-to-x16 risers in the as-built system.

2.2 Historical smallpieces v2 Slot Map (Octominer X12 Ultra — Jun 30, 2026 to Oct 25, 2026)

Prior to donating the Octominer X12 Ultra chassis to scraps v2 on October 26, 2026, smallpieces v2 mapped 6 GPUs across silk-screened Octominer motherboard slots (Slot 1: CPU-direct PCIe Gen3 x16; Slots 2–12: chipset PCIe Gen3 x1):

🐳 Container Runtime & Hardware Swap Procedure: See Fleet Docker Compose Architecture & Hardware Lifecycle for the standardized /opt/docker-configs/fah-compose.yaml specification, fah-start / fah-stop timer lifecycle, and graceful GPU swap procedure shared across all fleet hosts.

---

🚀 3. Hardware Upgrade Roadmap (Completed October 29, 2026)

1. Phase 1 (Slot Swap - Completed July 17, 2026): Relocated RTX 4070 from Slot 10 to physical Slot 1 (CPU-adjacent x16 slot) and moved MSI SHADOW RTX 5060 Ti to Slot 10 (x1 backplane slot). Verified system stability and unthrottled x16 link throughput (+1.49M PPD net pair surge).
2. Phase 2 (RTX 5070 Slot 1 Upgrade & 6-GPU Array Expansion - Completed July 18, 2026): Installed MSI GeForce RTX 5070 VENTUS 2X OC in physical Slot 1 (PCIe Gen3 x16), relocated ZOTAC RTX 4070 to Slot 6 (x1 backplane slot), and expanded live smallpieces array to 6 GPUs.
   • Initial Observations:
     • Slot 1 (x16) Live WebSockets & Overnight Steady-State Truth (gpu:01:00:00): Overnight telemetry across 12 completed Work Units establishes the steady-state performance ceiling for the RTX 5070 VENTUS in Slot 1 x16 (gpu:01:00:00 | UUID d87d48a0). On benchmark Project 18246, TPF converged down to 1m43.3s (103.3s), yielding a peak production rate of 9.01M PPD and a 95th percentile rate of 8.86M PPD (+9.3% net PPD surge and -6.2s/frame speedup over the RTX 4070 N-1 baseline in the same slot). High-frequency NVML profiling confirms peak PCIe bus throughput of 1.79 GB/s without link throttling.
     • Slot 6 (x1) 4070 Performance (gpu:04:00:00): RTX 4070 (gpu:04:00:00 in x1 backplane slot | UUID d65f3411) yields 4.66M PPD (@ 145W cap / 83.2W real load). High-frequency NVML sampler records peak burst transfers of 1.21 GB/s during CUDA frame step exchanges.
     • Array & Fleet Real-Time Yield: Live smallpieces v2 host output streamed at 27.60M PPD, bringing total live fleet production to ~66.5M PPD (projected ceiling ~69.2M PPD when itztli was not running).
     • Thermal & Power Stability: Total 6-GPU array power draw measured at 616.0W on 240V power with zero thermal throttling or watchdog faults.
3. Phase 3 (Blackwell Scheduling & 7-GPU Array Feasibility - July 19, 2026):
   • Ada Spin-Locking vs. Blackwell Event Yielding: Overnight high CPU loads and container watchdog resets (fah-runaway) were traced to Ada Lovelace (RTX 4070) driver behavior (cuEventSynchronize busy-wait spin-locking), which pegged CPU cores at 93.2%. Blackwell cards (cudaDeviceScheduleYield) yield CPU threads cleanly to kernel idle (intel_idle), dropping host CPU utilization to 71.5% with 28.5% idle headroom under 5 Blackwell GPUs.
   • Slot 1 GPU CPU Workload Progression: Empirical baseline telemetry across regimes confirms that higher-tier cards in CPU-tied Slot 1 x16 demand progressively greater CPU overhead to service host-to-device event polling and Hilbert grid reordering:
     • Regime N-2 (RTX 5060 Ti in Slot 1 x16): ~11.5% host CPU demand (total 5-GPU host load 65.8%).
     • Regime N-1 (RTX 4070 in Slot 1 x16): ~13.2% host CPU demand (total 5-GPU host load 68.4%).
     • Regime N (RTX 5070 in Slot 1 x16 + 6th GPU): ~15.5% host CPU demand (total 6-GPU host load 93.2%).
   • 7-Blackwell GPU Array Feasibility Model: Because Blackwell GPUs yield host CPU threads cleanly, scaling smallpieces to 7 Blackwell GPUs (1x 5070 Dual + 6x 5060 Ti) is physically possible on the Core i7-7700 host, projecting 36.95M PPD at 41.60 WUs / day (~88-92% CPU load).
   • CRITICAL 4070 EXCLUSION RULE: Running 7 GPUs is ONLY possible if ALL installed cards are Blackwell (RTX 50-series). Running 7 GPUs with the RTX 4070 present would push host CPU load past 105%, triggering severe thread lockups and watchdog resets (fah-runaway).
   • Platform Upgrade Conclusion: The Intel B250 / LGA 1151v1 chipset natively maxes out at 4C/8T processors (i7-7700). While 6 Blackwell GPUs (or 7 Blackwell GPUs with all-50-series cards) represent the physical ceiling for this chassis, further compute expansion beyond 7 cards will require a next-generation server architecture.
4. Phase 4 (smallpieces v3 5-GPU All-Blackwell Cutover Completed Oct 29, 2026 & scraps v2 Octominer Cutover Completed Oct 26, 2026):
   • RTX 5070 CPU Starvation & Pre-Shutdown 6→1 GPU Drain Benchmark (Oct 25, 2026): Prior to powering down smallpieces v2 for the v3 rebuild, all 6 GPUs were set to Finish while folding Project 18264 (WU1964–WU1969), creating a controlled step-down benchmark from 6 concurrent GPUs down to 1 solo GPU on the outgoing i7-7700 (4C/8T) + PCIe Gen 3.0 x1 Octominer platform:
     • WU1969 (RTX 5060 Ti, PCIe 3.0 x1, 150W cap): Frame time (TPF) dropped monotonically as sibling GPUs finished and vacated the bus/CPU: 309.67s (3.59M PPD) at 6 GPUs → 296.57s (3.83M PPD) at 5 GPUs → 278.79s (4.20M PPD) at 4 GPUs → 272.25s (4.36M PPD) at 3 GPUs → 261.22s (4.64M PPD) at 2 GPUs → 249.75s (4.96M PPD) solo at 1 GPU (-59.92s / -19.4% TPF, +38.1% PPD).
     • WU1968 (RTX 5070, Slot 1 PCIe 3.0 x16, 175W cap): TPF dropped from 193.93s (7.25M PPD) at 6 GPUs down to 176.50s (8.35M PPD, +15.2% PPD) at 4 GPUs even on an unthrottled x16 slot, isolating host CPU feeder contention (6x FahCore_27 competing for 4C/8T) from PCIe 3.0 x1 (~985 MB/s) link saturation.
   • The smallpieces v3 Unstarved Cutover (Oct 29, 2026 — Zen 5 Ryzen 9 9900X + PCIe Gen 5.0/4.0 x4 — 45.33M PPD on 5 GPUs): Following scraps v2's October 26 cutover into the vacated Octominer X12 Ultra chassis (i7-7700 @ 2.40 GHz sweet spot + 32GB DDR3L + 1x RTX 3070 + 3x RTX 3060 Ti), smallpieces v3 went live on October 29, 2026 inside the repurposed GPNE chassis (standing vertically on its end with ~2 PCI slot widths between cards) on a 12-core / 24-thread AMD Ryzen 9 9900X with 32GB (2x16GB) Kingston FURY DDR5-6000 RAM (running at 6000 MT/s) (paired with OPENMM_CPU_THREADS=4 in Docker to prevent 24-thread CPU reference pools from ballooning host RES memory) and unthrottled PCIe Gen 5.0 x4 (R34A Slots 1–4 via 2x 40cm + 2x 50cm ADT-Link K43SM risers) and native PCIe Gen 4.0 x4 (PCIEX16_2 direct on motherboard, no riser) lanes across all 5 Blackwell GPUs (1x RTX 5070 + 4x RTX 5060 Ti). The 5-GPU Blackwell setup surged to 45.33M PPD (RTX 5070: 13.01M PPD at 36.58s TPF on P18257, edging out origami's Zotac RTX 5070 at 37.08s / 12.45M PPD; 4x RTX 5060 Ti: 7.88M–8.27M PPD each)—surpassing the old 6-GPU smallpieces v2 Octominer output (27.60M PPD) by +64.2% while dropping host CPU utilization from 93.2% down to 22.2% and lifting mean GPU SM occupancy from 83.05% to 96.36%.

---

🚀 4. smallpieces v3 Architecture & Live Cutover Record (Completed October 29, 2026)

Cutover Status — LIVE (October 29, 2026): smallpieces v3 is live in the repurposed GPNE 8-fan wind-tunnel chassis (standing vertically on its end) on Ubuntu 26.04.1 LTS (Linux 7.0.0-42-generic, NVIDIA Driver 615.71.09), streaming 45.33M PPD across its pure 5-GPU Blackwell array (1x RTX 5070 + 4x RTX 5060 Ti) with 32GB (2x16GB) Kingston FURY DDR5-6000 (running at 6000 MT/s) RAM (OPENMM_CPU_THREADS=4), just 22.2% CPU utilization on the 12-core / 24-thread AMD Ryzen 9 9900X, and 96.36% mean GPU SM utilization.

4.1 Executive Summary & Architectural Overview

smallpieces v3 is a high-density 5-GPU All-Blackwell server engineered for high-throughput Folding@Home (F@H) work unit processing without PCIe Gen3 x1 backplane bottlenecks or 4-core CPU feeder starvation. Built inside the repurposed GPNE 8-fan wind-tunnel chassis (standing vertically on its end with ~2 PCI slot widths between cards) using a 12-core AMD Ryzen 9 9900X CPU, 32GB (2x16GB) Kingston FURY DDR5-6000 (running at 6000 MT/s) memory, and an ASUS TUF Gaming X670E-PLUS WiFi 6E motherboard, the build leverages native x4/x4/x4/x4 PCIe bifurcation on primary CPU slot PCIEX16_1 via a passive ADT-Link R34A quad M.2 adapter and 4x shielded ADT-Link K43SM M.2-to-PCIe x16 risers (2x 40cm and 2x 50cm)—all four linking at the motherboard's native PCIe Gen 5.0 x4 (32 GT/s)—plus direct motherboard seating on chipset slot PCIEX16_2 (no riser, linking at native PCIe Gen 4.0 x4 (16 GT/s)), two CPU-direct WD PC SN740 256GB NVMe SSDs, a 2400W single-rail 12V server PSU with a 12V Pico-PSU, and two Alphacool ES Guardian 4-Port Powerfan Hubs (CHA_FAN2 Intake & CHA_FAN1 Exhaust, regulated by gpufanpin v1.4.0). There are no x16-to-x16 nor x4-to-x16 risers in the as-built system.

4.2 System Bill of Materials (BoM)

Core Platform, Storage, Riser & Power Infrastructure

4.3 2400W Server PSU Power Distribution Topology (As-Built 10-Lead Allocation)

4.4 Operational Details

• OS, Kernel & Driver Stack: Ubuntu 26.04.1 LTS, Linux 7.0.0-42-generic, NVIDIA UNIX Open Kernel Module 615.71.09, Folding@Home v8.5.6.
• 5-GPU Production Breakout (45.3M–46.8M PPD vs 27.60M PPD on 6 GPUs in v2): Across 283+ credited WUs (Nov 2–12), the unstarved 5-GPU Blackwell array averaged ~60.5k PPD/W GPU efficiency (RTX 5070 d87d48a0 at 185W cap: 14.07M–14.73M PPD @ 76.0k–80.4k PPD/W, gaining +1.90% / +269k PPD @ 26.6k marginal PPD/dW over 175W and beating origami's EPYC 7543 ZOTAC RTX 5070 dd40d876 across 4/4 shared projects by +4.7% to +8.9%; 4x RTX 5060 Ti: 7.45M–8.35M median PPD each), outperforming the former 6-GPU smallpieces v2 Octominer array (27.60M PPD typical) by +64% to +70% with 1 fewer GPU.
• Power, GPU + Chassis Fan Regulation (gpufanpin v1.4.0) & Heat Comparison (smallpieces v3 vs scraps v2):
  • 5x Blackwell GPUs: 754.3W median live DC draw (~745–765W against 795W total nvpowerctl cap [185, 150, 150, 160, 150 W — recalibrated Nov 10 on GPU 0 175➔185W and GPU 3 16GB 150➔160W for +332k like-for-like PPD at +13.0W silicon / -0.75% wall PPD/W]; 59–67°C with gpufanpin.service v1.4.0 calling NVML's native nvmlDeviceSetFanSpeed_v2 / nvmlDeviceSetDefaultFanSpeed_v2 symbols in libnvidia-ml.so.1 via Python ctypes—zero Xorg/Coolbits dependency—to pin both GPU 0 [RTX 5070] and GPU 3 [PNY RTX 5060 Ti 16GB, bottom card] at 75% fan speed).
  • Dual Alphacool ES Guardian Chassis Hubs (CHA2_Intake & CHA1_Exhaust via nct6799) & 1" Audimute PSU Hush Box: Upgraded Nov 13, 2026 with two Alphacool ES Guardian 4-Port Powerfan Hubs (no USB header; pure 4-pin MB PWM + tach on Nuvoton nct6799, each module driving 4x M12038 120x38mm 1.50A fans and powered by two Molex leads off its own 6-pin PSU lead [Leads 1 & 9]), paired with a custom 1" Audimute Eco-C-Tex acoustic hush box over the LX2400W server PSU bay to attenuate its shrill internal fan. Because the LX2400W outputs continuous +12V whenever its AC rocker is ON, both hubs default to 100% open-collector PWM failsafe whenever the motherboard is unpowered (S5 soft-off or pre-POST), mirroring Octominer failsafe behavior. Once booted, gpufanpin v1.4.0 declaratively syncs /etc/sensors.d/gpufanpin-sensors.conf (CHA1_Exhaust, CPU_FAN, CHA2_Intake, CPU_Opt, MB_SYSTIN, MB_CPUTIN, CPU_PECI) and drives RPM-balanced raw 0–255 PWM curves (chassis_idle_pwm = 4, 0 ➔ ~850–960 / 1,520 RPM idle; chassis_load_min_pwm = 26, 6 ➔ chassis_load_max_pwm = 102, 71 across 48°C–74°C max GPU temp), balancing steady-state 5-GPU datacloset F@H load around ~2,250–2,350 RPM Intake / ~2,350–2,420 RPM Exhaust with proportional ceiling headroom up to ~2,650 RPM at 74°C (with a 255 emergency latch at >=78°C GPU / >=90°C CPU).
  • Ryzen 9 9900X CPU: 88.0W RAPL package power (pinned at the 65W Eco Mode 88W PPT ceiling under performance governor; 65–67°C at 22.2% thread busy across 24 threads) vs 18.2W RAPL package power on scraps v2's i7-7700 @ 3.00 GHz 5-GPU sweet spot.
  • Total Thermal Output: Including motherboard/DDR5/NVMe/8x GPNE fans (~46W DC) and 240V PSU conversion loss (~85W), smallpieces v3 dissipates ~890W DC / ~965–980W AC wall (~3,290–3,345 BTU/hr).
• Host RAM Relief via OPENMM_CPU_THREADS=4 (25.95 GiB ➔ ~14 GiB on 32GB Kingston FURY DDR5-6000): Because FahCore_27 initializes an un-parameterized 3-argument OpenMM::Context("CPU") reference instance per GPU (used strictly once at every 5% checkpoint to sanity-check a single frame of GPU forces), moving from the 8-thread i7-7700 to the 24-thread Ryzen 9 9900X caused OpenMM to spawn 104 threads/GPU (520 total across 5 GPUs) and pre-allocate 4.45–6.29 GiB/GPU (25.95 GiB total host RSS on the 32GB Kingston FURY DDR5-6000 host) in idle per-thread PME 3D FFT grid and force buffers. Setting OPENMM_CPU_THREADS=4 in /opt/docker-configs/fah-compose.yaml caps the CPU sanity-check thread pools at 4 threads (24 threads/GPU, 120 total), reclaiming ~11–12.5 GiB of host DDR5 RAM (dropping 5-GPU FahCore_27 RSS to ~13.5–15.0 GiB) with zero impact on GPU TPF or PPD.

4.5 Step-by-Step Assembly, Fresh OS Installation & NAS Restore Record

1. Step 1: Decommission smallpieces v2 (to Build scraps v2) & Prepare GPNE Chassis
   • Decommission smallpieces v2 by pulling its mSATA + 2.5" SATA drives and Blackwell GPUs from the Octominer X12 Ultra chassis, and transplant scraps v1's drives and 4 Ampere GPUs (1x RTX 3070 + 3x RTX 3060 Ti) into the Octominer X12 Ultra to produce scraps v2.
   • Unscrew and remove the retired i7-4790 AX-B85M-ETH riserless mining board from the GPNE 8-fan wind-tunnel chassis.
   • CRITICAL ELECTRICAL SAFETY: Remove all non-standard metal standoffs from the GPNE chassis tray. Install standoffs strictly matching ATX motherboard holes. Place insulating electrical tape over any non-removable metal nubs underneath the motherboard area.
2. Step 2: Core Hardware & Power Harness Installation
   • Install the AMD Ryzen 9 9900X CPU into the ASUS TUF Gaming X670E-PLUS WiFi AM5 socket. Apply thermal compound and mount the Thermalright AXP90-X53 Full Copper cooler.
   • Install 32GB (2x16GB) Kingston FURY DDR5-6000 RAM modules (running at 6000 MT/s) into dual-channel slots (DIMM_A2 + DIMM_B2).
   • Plug the 200W–250W Pico-PSU module into the 24-pin ATX header on Lead 5, leaving the Pico-PSU's 8-pin ATX CPU pigtail unplugged and velcroed off. Connect the ModDIY 6-pin PSU to 8-pin ATX CPU pigtail on Lead 6 directly to motherboard ATX_12V_1 (ATX_12V_2 is not connected), connect the two 6-pin PSU to 3x SATA strings to Leads 7 & 8, connect Leads 1 & 9 (6-pin PSU to Dual Molex) to the two Alphacool ES Guardian 4-Port Powerfan Hubs (Controller 1 Intake ➔ CHA_FAN2, Controller 2 Exhaust ➔ CHA_FAN1), and velcro off unused Lead 10.
   • Secure the ASUS TUF Gaming X670E-PLUS WiFi motherboard to the GPNE chassis floor.
3. Step 3: Expansion, Storage, Clean Ubuntu 26.04.1 LTS Install & LAN NAS /backup Restore
   • Insert the passive ADT-Link R34A (PCIe 5.0 x16 to Quad M.2 NVMe bifurcation card) into primary CPU slot PCIEX16_1 (leave its optional 2-pin fan header unconnected).
   • Install the 4x ADT-Link K43SM (2x 40cm + 2x 50cm) M.2-to-PCIe x16 risers into R34A Slots 1–4 (Slot 1: MSI RTX 5070 VENTUS 2X OC d87d48a0 on Lead 2, Slot 2: MSI RTX 5060 Ti VENTUS 56d10701 on Lead 4, Slot 3: MSI RTX 5060 Ti SHADOW #1 a83e55f9 on Lead 3, Slot 4: PNY RTX 5060 Ti 16GB VERTO 2e4a1d96 on Lead 3), and plug the 5th GPU (MSI RTX 5060 Ti SHADOW #2 8854baa9 on Lead 4) directly into motherboard slot PCIEX16_2 (no riser). Leave bottom-edge slot PCIEX4 unpopulated.
   • Install the two WD PC SN740 256GB NVMe 1.4 SSDs into the two CPU-direct M.2 slots: Storage #1 (BOOT/OS) into M.2_1 (/dev/nvme0n1, Key M, 2280, PCIe 5.0 x4) and Storage #2 (DATA/OPT) into M.2_3 (/dev/nvme1n1, Key M, 2280, PCIe 4.0 x4), leaving chipset slots M.2_2 and M.2_4 empty.
   • Fresh Ubuntu 26.04.1 LTS Installation (Built from Scratch): Install a fresh copy of Ubuntu 26.04.1 LTS (Linux 7.0.0-42-generic) directly onto Storage #1 (M.2_1, /dev/nvme0n1 mounted at /) with Storage #2 (M.2_3, /dev/nvme1n1) mounted at /opt in /etc/fstab (by UUID), and install baseline packages (NVIDIA Open Kernel Module driver 615.71.09, docker.io, nvidia-container-toolkit, lm-sensors).
   • Restore from /backup off the LAN NAS: Mount /backup from the LAN's NAS and restore smallpieces's user/group accounts, SSH host and user keys, /home/, the /opt/{bin,etc,logs,docker-configs,src} tree, cron/logrotate configurations, and custom systemd services (omitting octoctl and intel-undervolt, which remain on the Octominer X12 Ultra running scraps v2).
4. Step 4: ASUS TUF Gaming X670E-PLUS WiFi BIOS & Post-Cutover OS Setup
   • Power on and press DEL (then F7 for Advanced Mode) to enter ASUS UEFI.
   • Slot Bifurcation (x4/x4/x4/x4) & Link Speed (Auto): Navigate to Advanced > Onboard Devices Configuration and set PCIEX16_1 Bandwidth Bifurcation Configuration to PCIe RAID Mode (x4/x4/x4/x4), ensure NVMe RAID Mode under Advanced > AMD PBS is Disabled, and leave both PCIEX16_1 Link Speed and PCIEX16_2 Link Speed on Auto (allowing the 4 R34A K43SM risers on PCIEX16_1 to negotiate at full PCIe Gen 5.0 x4 / 32 GT/s and PCIEX16_2 to negotiate cleanly at PCIe Gen 4.0 x4 / 16 GT/s; manually forcing Gen4 on PCIEX16_2 causes X670E chipset downstream port 0b:00.0 link training against the native Gen 5 Blackwell GPU to fail).
   • Above 4G Decoding = Enabled (41bit / 2TB) & Re-Size BAR Support = Disabled: Under Advanced > PCI Subsystem Settings, set Above 4G Decoding to Enabled, set Above 4G MMIO Limit to 41bit (2TB), and set Re-Size BAR Support to Disabled (paired with pci=realloc,noaer pcie_aspm=off iommu=pt in /etc/default/grub). Leaving Re-Size BAR enabled across the CPU-direct GPUs causes each card to claim 16–32 GB of MMIO bridge window space during POST, exhausting the X670E chipset bridge (00:02.1) aperture and starving both PCIEX16_2 and the onboard Realtek 2.5GbE NIC (eno1) with can't claim; no compatible bridge window.
   • AMD 65W Eco Mode (88W PPT) & Qfan: Under Ai Tweaker > Precision Boost Overdrive (or Advanced > AMD Overclocking > Precision Boost Overdrive), enable Eco Mode (65W / 88W PPT) to match the Thermalright AXP90-X53 Full Copper cooler, and set CPU_FAN in Qfan (F6) to Turbo / Full Speed.
   • Boot into Linux (hostname smallpieces), verify all 5 Blackwell GPUs link at x4 width (4x PCIe Gen5 x4 on PCIEX16_1 R34A Slots 1–4 + 1x native PCIe Gen4 x4 on PCIEX16_2), verify GPU index order with nvidia-smi -L, and confirm /opt/etc/nvpowerctl.conf maps 185W to the RTX 5070, 160W to the 16GB RTX 5060 Ti, and 150W to the three 8GB RTX 5060 Ti cards (gpu_limits = 185, 150, 150, 160, 150).