← Back to Fleet Dashboard

🗑️ scraps v2 (Octominer X12 Ultra)

🖥️ Hardware Summary (scraps v2 — Live Since Oct 26, 2026; 5-GPU Expansion Nov 1, 2026)

• Chassis & Cooling: Octominer X12 Ultra 12-Slot Server Chassis with high-static-pressure Delta fan wall (closed-loop PWM + CPU thermal-assist control via octoctl v2.2.15) & Thermalright AXP90-X36 CPU cooler
• Motherboard & CPU: Octominer B250 Custom Board + Intel Core i7-7700 (Kaby Lake 4-Core / 8-Thread; unlocked via nvpowerctl v1.3.4 clearing MSR 0x1a0 bit 38 with 3.00 GHz sweet-spot ceiling & per-thread FahCore CPU affinity isolation)
• System Memory (RAM): 32GB DDR3L-1600 SODIMM (2x 16GB, OPENMM_CPU_THREADS=4) — 4x upgrade over v1's 8GB ceiling (0.00 KB/s swap)
• Network Interface: Onboard Realtek RTL8105E Fast Ethernet (100baseT/Full, 100 Mbps / ~11.2 MB/s wire speed on enp3s0)
• Power Infrastructure: 4x HPE 750W Platinum Plus 240V Server PSUs (3000W total capacity, direct riserless motherboard busbars)
• OS, Kernel & Driver: Ubuntu 26.04.1 LTS | Linux Kernel 7.0.0-42-generic | NVIDIA Driver 595.91.07 (CUDA 13.2)
• Current GPUs (5 Active Ada/Ampere GPUs — 20.20M PPD Nov 1 baseline; tuned Nov 10 to 725W cap / 674.5W real draw for +472k like-for-like PPD [~20.6M–21.5M PPD] & 52°C RTX 3070): 1x ZOTAC GeForce RTX 4070 Blower (12GB, Slot 1 Gen3 x16, 160W cap, 65% fan floor via gpufanpin v1.3.1) + 1x ZOTAC GeForce RTX 3070 (8GB, Slot 2 Gen3 x1, 145W cap, 65% fan floor via gpufanpin v1.3.1) + 3x GeForce RTX 3060 Ti (8GB, Slots 4, 7, 10 Gen3 x1, 140W cap) — 725W combined GPU power cap (160, 140, 145, 140, 140)
• Retired Hardware (scraps v1): Intel Core i7-4790 + 8GB DDR3 AX-B85M-ETH board, 1x ASUS RTX 2060 (6GB, UUID 535ba65c), and 1x Gigabyte RTX 2060 SUPER (8GB, UUID 8990101d)

🗺️ Physical to Logical GPU Ordering Map (Octominer X12 Ultra — scraps v2 5-GPU Array)

The physical slot designations are silk-screened on the Octominer X12 Ultra motherboard. Slot 1 is wired directly to the CPU as an unthrottled PCIe Gen 3.0 x16 link (housing the ZOTAC GeForce RTX 4070 Blower added November 1, 2026); Slots 2–12 are riserless physical x16 slots wired to the B250 PCH as PCIe Gen 3.0 x1 links (~985 MB/s per slot, housing the ZOTAC GeForce RTX 3070 in Slot 2 and the three RTX 3060 Ti cards in Slots 4, 7, and 10).

🐳 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.

---

🚀 Scraps v2 Upgrade: Octominer X12 Ultra Migration (Oct 26, 2026) & 5-GPU RTX 4070 Expansion (Nov 1, 2026)

Cutover Status — LIVE (Oct 26, 2026; Expanded to 5 GPUs Nov 1, 2026; QRB Power-Cap Tuned Nov 10, 2026): scraps v2 is live inside the Octominer X12 Ultra 12-slot server chassis at its original IP (.27), running its migrated mSATA/SATA storage volumes intact, closed-loop octoctl v2.2.15 fan control, gpufanpin v1.3.1 (65% floor on both GPU 0 RTX 4070 and GPU 2 RTX 3070), nvpowerctl v1.3.4 (160W, 140W, 145W, 140W, 140W GPU caps = 725W total [raised Nov 10 from 665W to lift the 3x RTX 3060 Ti off their V_min voltage floor by +120..+255 MHz and RTX 4070 to 2,550 MHz] + MSR 0x1a0 unlock + 3.00 GHz ceiling + fah_main_cpus = 0, 3, 1, 2, 7), and its 5-GPU Ada/Ampere array (1x RTX 4070 + 1x RTX 3070 + 3x RTX 3060 Ti). Meanwhile, smallpieces v3 (AMD Ryzen 9 9900X + PCIe Gen 5.0/4.0 x4 in the repurposed GPNE chassis) is live with its 5-GPU all-Blackwell array.

1. Executive Summary & Fleet Hardware Cascade

The legacy scraps v1 platform (Haswell Core i7-4790 on an AX-B85M-ETH board) was constrained by an immovable 8GB single-SODIMM DDR3 ceiling, PCIe 2.0 x1 slots, and a shared 120V / 15A circuit. Rather than purchasing a separate motherboard/CPU kit for scraps, the fleet executed a hardware cascade across smallpieces and scraps:

1. Stage 1 (scraps v2 Migration into Octominer X12 Ultra — Completed Oct 26, 2026): Transplanted scraps's existing 128GB mSATA boot drive and 30GB 2.5" SATA /opt drive into the vacated Octominer X12 Ultra chassis (Core i7-7700 4C/8T, 32GB DDR3L RAM, Thermalright AXP90-X36 cooler, and 4x HPE 750W 240V Platinum Plus PSUs), enabled octoctl.service, unlocked the B250 BIOS 800 MHz lock via nvpowerctl v1.3.2, and initially deployed 4 Ampere GPUs (1x Zotac RTX 3070 in Slot 1 + 3x RTX 3060 Ti in Slots 4, 7, 10).
2. Stage 2 (smallpieces v3 All-Blackwell Cutover — Completed Oct 29, 2026): Rebuilt smallpieces v3 (Ryzen 9 9900X + ASUS TUF Gaming X670E-PLUS WiFi 6E + 32GB Kingston FURY DDR5-6000 + ADT-Link R34A Gen5 x4 bifurcation + 2400W PSU) inside the repurposed GPNE frame for smallpieces's 5 Blackwell GPUs (1x RTX 5070 + 4x RTX 5060 Ti).
3. Stage 3 (scraps v2 5-GPU Expansion with ZOTAC RTX 4070 Blower — Completed Nov 1, 2026): Moved the ZOTAC GeForce RTX 3070 (1522578d, 145W) from OX12U Slot 1 to OX12U Slot 2 (Gen3 x1, Bus 04:00.0, GPU 2) and installed the ZOTAC GeForce RTX 4070 Blower (d65f3411, 12GB GDDR6X, 145W) into CPU-direct OX12U Slot 1 (Gen3 x16, Bus 01:00.0, GPU 0) alongside the three RTX 3060 Ti cards in Slots 4, 7, and 10.

2. Core Platform & GPU Lineup Comparison (scraps v1 ➔ scraps v2)

---

📊 Live Per-GPU Benchmarks: scraps v1 vs. scraps v2 (4-GPU & 5-GPU Regimes)

Shortly after initial cutover on October 26, 2026, all four Ampere GPUs in scraps v2 drew concurrent Project 15506 (FahCore_28 v0.1.3, 166,465 atoms) Work Units from Temple University's assignment server (vav17.fah.temple.edu), enabling a direct apples-to-apples comparison against historical Project 15506 runs on the exact same physical RTX 3060 Ti silicon inside scraps v1 (i7-4790 + PCIe Gen 2 x1). Following the November 1, 2026 expansion to 5 GPUs (adding the ZOTAC GeForce RTX 4070 Blower into CPU-direct Slot 1 Gen3 x16 and moving the ZOTAC GeForce RTX 3070 to Slot 2 PCH Gen3 x1 under 3.00 GHz + nvpowerctl v1.3.4 thread affinity), we measured the full 5-GPU lineup across Project 15506 (FahCore_28) and Project 18249 (FahCore_27):

• Why Slot 1 (CPU-Direct Gen3 x16) Belongs to the RTX 4070: Moving the ZOTAC RTX 3070 from Slot 1 (Gen3 x16) to Slot 2 (PCH Gen3 x1) cost +2.5–3.0s TPF on Project 15506 (43.52s ➔ 46.5s, or -0.59M PPD), but freed CPU-direct Slot 1 (15.8 GB/s) for the ZOTAC RTX 4070 Blower (d65f3411), which bursts up to 2,997.64 MB/s TX / 1,132.39 MB/s RX on Project 18249 and adds +6.05M–6.62M PPD (41.7k–45.6k PPD/W) at the same 145W cap—yielding a net node gain of +6.50M PPD (13.70M ➔ 20.20M PPD).
• 5-GPU Node Production Summary (20.20M PPD @ 647.3W): At the 3.00 GHz (0.980V, 18.5W) 5-GPU operating point with nvpowerctl v1.3.4 thread affinity (fah_main_cpus = 0, 3, 1, 2, 7), scraps v2 sustains 20.20M PPD (6.05M–6.62M + 4.10M + 3 × ~3.30M–3.52M) at 628.8W combined GPU power and 18.5W CPU package power (647.3W total silicon draw = 31.20k PPD/Watt).

Whole-Node Production & Silicon Efficiency: scraps v2 (4-GPU & 5-GPU Regimes) vs. scraps v1 Historical Telemetry (57,641 Samples)

Comparing 57,641 full-load 1-minute telemetry samples from scraps v1's historical neumon.log across both of its operational eras—Phase A (Jul 25 – Aug 28, 2026: 5 GPUs) and Phase B (Aug 29 – Oct 25, 2026: 3 GPUs + 2 idled Turings)—against both stages of scraps v2 (Oct 26–31 4-GPU Cutover and Nov 1+ 5-GPU Expansion) shows how replacing the Haswell/GPNE platform and Turing cards doubled output and efficiency at the exact same wall power:

• Doubling 5-GPU scraps v1 Output (10.11M ➔ 20.20M PPD) for Just +10.4W (15.87k ➔ 31.20k PPD/Watt): Replacing the two Turing 20-series cards (RTX 2060 SUPER + RTX 2060) with the ZOTAC RTX 4070 Blower (145W) and ZOTAC RTX 3070 (145W) while dropping CPU package power from 50.8–60.4W (i7-4790) to 18.5W (i7-7700 @ 3.00 GHz) delivers literally 2.0x the PPD (20.20M vs. 10.11M) at within 1.6% (+10.4W) of the exact same total silicon power—lifting whole-node efficiency by +96.6% (31.20k vs. 15.87k PPD/W) and tripling (3.09x) the 6.53M PPD output of 3-GPU scraps v1 Phase B.
• 4-GPU Intermediate Baseline (Oct 26–31: 13.70M–14.61M PPD @ 519.2W): During its first five days with 4 Ampere GPUs (1x RTX 3070 + 3x RTX 3060 Ti at 2.40 GHz / 10.4W CPU), scraps v2 sustained 13.70M–14.61M PPD (43.52s RTX 3070 TPF and 54.58–55.51s 3060 Ti TPF on P15506) at 26.39k PPD/W—already beating 5-GPU scraps v1 by +35.5% while burning 117.7W less power.

How LAR.systems Rankings Work & Why scraps v2 Scores "Unicorn (+88%)" on P18264 vs "Normal" / "OK" on Un-Skewed Desktop Baselines

LAR.systems computes its slot performance badges (UNICORN >= +30%, GREAT +20..30%, GOOD +10..20%, NORMAL -10..+10%, OK -20..-10%, POOR -30..-20%, OOF < -30%) strictly per-(GPU Model, OS, Project ID) via model-specific JSON profiles (ad104_geforce_rtx_4070.json, ga104_geforce_rtx_3070.json, etc.). Because the LAR webclient periodically pushes active slot samples from open browser tabs, comparing the ZOTAC RTX 4070 Blower (d65f3411, 145W cap) across its first two Alzheimer's projects (919,221 atoms) on scraps v2 reveals two distinct phenomena:

• Self-Referential "Unicorn (+88%)" on Project 18264 (6.21M PPD vs LAR's 3.31M PPD Linux Average across 1,112 samples): Nearly all 1,112 Linux P18264 samples in LAR's ad104_geforce_rtx_4070.json database were submitted by this exact RTX 4070 during the smallpieces v2 era—when it was choked at ~3.3M PPD on a PCIe Gen 3.0 x1 riser (Slot 6) with the i7-7700 BIOS-locked at 800 MHz (MSR 0x1a0 bit 38). Moving the card into scraps v2 Slot 1 (PCIe Gen 3.0 x16) with the i7-7700 unlocked to 3.00 GHz and TID == PID thread affinity isolated lifted its P18264 speed by +88.0% (3.31M ➔ 6.21M PPD), earning a live "UNICORN" badge against its own historical baseline!
• "Normal" / "OK" (-6% to -14%) on Un-Skewed Desktop Baselines (P18258, P15506): On projects where LAR's Linux average was set by un-capped 200–220W single-GPU desktop PCs on modern DDR4/DDR5 CPUs—such as P18258 on the RTX 4070 (~6.88M PPD vs LAR's 8.00M PPD average across n=15 samples = -14.0% "OK") or P15506 on the RTX 3070 (~4.10M PPD vs LAR's 4.36M PPD average = -6.0% "NORMAL")—five architectural differences between smallpieces v3 (Ryzen 9 9900X) and scraps v2 (i7-7700 Octominer) govern the gap:

---

🌡️ Kaby Lake i7-7700 Multiplier, Vcore & Thermal Sweet-Spot Analysis (2.40 GHz 4-GPU & 3.00 GHz 5-GPU Regimes)

When all GPUs initially began folding with MSR 0x1a0 bit 38 cleared and scaling_max_freq uncapped, two hardware behaviors of the Octominer B250 mining motherboard and Thermalright AXP90-X36 1U cooler surfaced:

1. Why smallpieces v2 Ran So Cool in the Same Chassis (50–56°C): Auditing 15,720 historical neumon.log samples from smallpieces v2 revealed that because the Octominer B250 BIOS boots with IA32_MISC_ENABLE (MSR 0x1a0) bit 38 set (0x4000850089), intel_pstate was locked into its minimum 8x (800 MHz) LFM state 93.1% of the time. At 800 MHz, the i7-7700 requests only 0.672 V Vcore and draws just 4.3–5.7 W package power—allowing the CPU to sit at 50–56°C even with the Octominer chassis fans idling at 22% (1,805 RPM).
2. Hard B250 BIOS Multiplier Ceiling at 36x (3.60 GHz) & Quadratic Vcore Heating: Even after clearing bit 38 of MSR 0x1a0 (0x850089), the Octominer B250 mining BIOS programs MSR_CONFIG_TDP_NOMINAL (0x648 = 0x24) and MSR_PERF_STATUS (0x198 = 0x2400) to cap the physical hardware multiplier at 36x (3.60 GHz base clock) regardless of MSR_TURBO_RATIO_LIMIT (0x1ad = 0x2829292a). Because each FahCore process busy-polls cuEventSynchronize continuously on one CPU thread, running 4–5 active GPU workers at 36x (3.60 GHz) drives Vcore to 1.14 V and CPU package power to 25.3–28.0 W (6x higher heat dissipation than 800 MHz), pushing the low-profile 36mm CPU cooler to 83–89°C unless the loud Delta chassis fans ramp to 85% (4,700 RPM).

Live Multiplier, Vcore, Package Power & TPF Sweep (4-GPU & 5-GPU Load)

By sweeping /sys/devices/system/cpu/cpufreq/policy*/scaling_max_freq across every Kaby Lake P-state multiplier under live 4-GPU and 5-GPU load, we mapped the non-linear V/F curve of the i7-7700 inside the Octominer X12 Ultra:

Production Daemon Configuration (nvpowerctl v1.3.4 + gpufanpin v1.3.1 + octoctl v2.2.15)

• nvpowerctl v1.3.4 (/opt/etc/nvpowerctl.conf): With the November 1, 2026 expansion to 5 GPUs (20.20M PPD @ 665W cap baseline) and the November 10, 2026 whole-system QRB power-cap recalibration to 725W total GPU cap (gpu_0 = 160, gpu_1 = 140, gpu_2 = 145, gpu_3 = 140, gpu_4 = 140 — un-starving the three RTX 3060 Ti cards from 1,380–1,575 MHz to 1,635–1,830 MHz and lifting the RTX 4070 from 2,445 ➔ 2,550 MHz), nvpowerctl v1.3.4 enforces cpu_turbo = enabled, cpu_scaling_governor = performance, cpu_max_freq_khz = 3000000 (3.00 GHz at ~0.98V / ~18.5W / ~66°C CPU package temp—keeping the i7-7700 9°C below octoctl's 75°C chassis fan ramp threshold while preventing idle logical cores from dropping to 800 MHz between OpenMM barrier wakes), and fah_thread_affinity = enabled (fah_main_cpus = 0, 3, 1, 2, 7):
  • How nvpowerctl Enforces CPU Affinity Under the Hood (apply_fah_thread_affinity): Each FahCore process spawns 28 threads in /proc/<pid>/task/<tid>: 1 main CUDA dispatch thread (TID == PID) that busy-polls cuEventSynchronize / ioctl at ~98–100% CPU (RNl), plus 27 helper threads (TID != PID) including 4 OpenMM CPU barrier threads (OPENMM_CPU_THREADS=4). On startup and every 15s watchdog poll, nvpowerctl scans /proc/[0-9]*/comm for FahCore*, extracts the GPU index from -gpu <idx> in /proc/<pid>/cmdline, and checks os.sched_getaffinity(pid) (idempotent—zero syscalls or log noise once pinned). Whenever a new Work Unit spawns a fresh FahCore PID, nvpowerctl iterates /proc/<pid>/task/* and calls Linux sched_setaffinity(2) directly via Python:
    1. Main CUDA Loop (tid == pid) ➔ Dedicated Single Logical Core ({target_cpu}): Pins GPU 0 (RTX 4070) to cpu0 (Core 0 primary), GPU 2 (RTX 3070) to cpu1 (Core 1 primary), GPU 3 (RTX 3060 Ti) to cpu2 (Core 2 primary), and GPU 1 & GPU 4 (2x RTX 3060 Ti) to cpu3 & cpu7 (Core 3's two hardware SMT threads)—guaranteeing two 100%-spinning main CUDA loops never land on the same logical core and time-slice each other 50%/50%.
    2. OpenMM Helper Threads (tid != pid) ➔ All Logical Cores ({0..7}): Keeps all 8 logical cores open for the 4 OpenMM barrier threads so step-boundary reductions execute in parallel across the unshared SMT siblings (cpu4, cpu5, cpu6) without serializing on a single core.
• gpufanpin v1.3.1 (/opt/etc/gpufanpin.conf): Pins both the ZOTAC GeForce RTX 4070 Blower (GPU 0, OX12U Slot 1) and the ZOTAC GeForce RTX 3070 (GPU 2, OX12U Slot 2; gpu_index = 0, 2, target_fan_speed = 65) to 65% via direct headless libnvidia-ml.so.1 C-library ctypes (nvmlDeviceSetFanSpeed_v2 / nvmlDeviceSetDefaultFanSpeed_v2 over /dev/nvidiactl, zero Xorg/Coolbits) during active folding, holding the RTX 4070 at 64–66°C at 160W cap and dropping the RTX 3070 median core temp from 60.0°C ➔ 52.0°C (-8.0°C, -0.9W leakage, -60% transient GA104 hotspot blips).
• octoctl v2.2.15 (/opt/etc/octoctl.conf):
  • Direct Super I/O CPU Cooler Fan Curve (nct6779 pwm2): Because the Nuvoton nct6795D Super I/O chip's autonomous SmartFan IV curve reads a sluggish PECI sensor (~15–20°C below true Intel coretemp Package ID 0), octoctl manages pwm2 directly using real coretemp package temperatures, ramping the Thermalright AXP90-X36 fan smoothly to 100% (~2,740 RPM) whenever the CPU reaches 65°C.
  • Secondary CPU Thermal Assist on Chassis Fan Wall (CPU_TEMP_START=75, CPU_TEMP_MAX=90): The 4 high-static-pressure Octominer Delta chassis fans stay on their quiet GPU curve (1,805–2,700 RPM when GPUs are 55–67°C) during 3.00 GHz operation (66°C CPU), and only ramp to assist the CPU cooler if CPU package temperature ever exceeds 75°C.

---

🔬 baseliner v3.7.1 Multi-GPU Scaling & Cycle Decomposition (0-GPU ➔ 1-GPU ➔ 2-GPU ➔ 4-GPU ➔ 5-GPU)

Using baseliner v3.7.1 (baseliner --compare), we captured 30-second high-frequency (250ms PCIe / NVML polling + hardware PMU cycle counters) baselines across all five operational load stages on scraps v2:

1. 0-GPU Idle Baseline (scraps_baseline_20261026_172421.log): All GPUs idle (P8 state, 34.5W total GPU power, CPU idling at 800 MHz).
2. 1-GPU Folding Baseline (scraps_baseline_20261026_175735.log): Slot 1 Zotac RTX 3070 folding P15506 (WU1823) solo at 3.60 GHz CPU clock.
3. 2-GPU Folding Baseline (scraps_baseline_20261026_182358.log): Slot 1 RTX 3070 + Slot 7 ASUS RTX 3060 Ti (WU1825) folding concurrently at 3.60 GHz CPU clock.
4. 4-GPU Full-Tilt Baseline @ 2.40 GHz (scraps_baseline_20261026_184741.log): All 4 Ampere GPUs (1x RTX 3070 + 3x RTX 3060 Ti) folding P15506 concurrently at the 2.40 GHz (0.864V, 10.4W, powersave governor) 4-GPU sweet spot.
5. 5-GPU Full-Tilt Production Baseline @ 3.00 GHz + Thread Affinity (scraps_baseline_20261101_161610.log): All 5 Ada/Ampere GPUs (GPU 0: RTX 4070 on Slot 1 Gen3 x16 + GPU 2: RTX 3070 on Slot 2 Gen3 x1 + GPU 1, 3, 4: 3x RTX 3060 Ti on Slots 10, 7, 4 Gen3 x1) folding concurrently (20.20M PPD) at the 3.00 GHz (0.980V, 18.5W, performance governor + fah_main_cpus = 0, 3, 1, 2, 7) 5-GPU production sweet spot.

Key Architectural Takeaways from the 5-Stage baseliner Progression

• 2,997.64 MB/s (3.00 GB/s) Peak PCIe Bursts Prove Why the RTX 4070 Needed Slot 1 (Gen3 x16): High-frequency 250ms PCIe telemetry in the 5-GPU capture (scraps_baseline_20261101_161610.log) caught GPU 0 (ZOTAC RTX 4070 Blower running FahCore_27 Project 18249 in Slot 1 CPU-direct Gen3 x16) bursting to 2,997.64 MB/s Peak TX and 1,132.39 MB/s Peak RX (190.14 MB/s avg TX / 61.61 MB/s avg RX)—3.0x the physical 985 MB/s ceiling of a B250 PCH PCIe Gen 3.0 x1 slot. Meanwhile, on the four PCH Gen 3.0 x1 slots, GPU 3 (ASUS RTX 3060 Ti running FahCore_27 P18249 in Slot 7) peaked at 908.57 MB/s TX / 803.47 MB/s RX (92% of its x1 link limit), while the three FahCore_28 (P15506) cards (GPU 1, 2, 4) peaked at 339–357 MB/s RX (35% of x1 link capacity).
• -71.5% Drop in Kernel %sys Time (18.6% ➔ 5.3%) & Fewer Context Switches with nvpowerctl v1.3.4 Thread Affinity: Even after adding a 5th GPU (raising hardware interrupts by +18.0% from 43,878/s to 51,796/s), moving from un-pinned powersave (2.40 GHz) to pinned fah_thread_affinity = 0, 3, 1, 2, 7 + performance (3.00 GHz) slashed kernel %sys CPU overhead from 18.6% down to 5.3% and reduced context switches by -7.3% (3,181/s ➔ 2,949/s), because isolating each FahCore main thread (TID == PID) onto its own logical core stops the Linux scheduler from migrating spinning CUDA loops across cores or time-slicing two main loops on the same runqueue.
• Onboard 100 Mbps Fast Ethernet & Work Server Sequential Downloads: The Octominer B250 (B8PLUS2.0) motherboard integrates a Realtek RTL8105E Fast Ethernet controller (100baseT/Full, ~11.2 MB/s max TCP payload throughput). When multiple GPUs request new Work Units simultaneously in the same F@H resource group, the F@H v8.4.9 client downloads multi-megabyte assignment payloads sequentially one-by-one (~8 seconds per 67.3 MiB P15506 WU over the 100 Mbps link once the upstream assignment server begins streaming).

---

3. Cutover Execution Notes & Historical scraps v1 Slot Map

• Network Interface Normalization: Upon first boot inside the Octominer B250 motherboard, the onboard Realtek RTL8105E NIC enumerated as enp3s0; updating Netplan restored scraps at its canonical static IP (10.42.42.27).
• Historical scraps v1 (AX-B85M-ETH) Slot Map (Jul 25, 2026 – Oct 25, 2026):
  • GPNE Slot 1 (x1, Bus 02): DELL RTX 3060 Ti (1ed25c7d) → Migrated to OX12U Slot 4 (Bus 06:00.0, GPU 4)
  • GPNE Slot 2 (x1, Bus 03): ASUS RTX 3060 Ti (a76d84a0) → Migrated to OX12U Slot 7 (Bus 05:00.0, GPU 3)
  • GPNE Slot 4 (x1, Bus 04): ASUS RTX 3060 Ti (d1ea526f) → Migrated to OX12U Slot 10 (Bus 02:00.0, GPU 1)
  • GPNE Slot 6 (x1, Bus 05): ASUS RTX 2060 (535ba65c, 6GB) → Idled Aug 29, 2026; Retired Oct 26, 2026
  • GPNE Slot 8 (x1, Bus 06): Gigabyte RTX 2060 SUPER (8990101d, 8GB) → Idled Aug 29, 2026; Retired Oct 26, 2026