← Back to home

origami v2 Hardware and Architecture

Platform Version: v2.0 (Active)
Architecture: AMD EPYC SP3 Server (128 PCIe 4.0 Lanes)
Memory: 256GB 8-Channel ECC RDIMM
Live GPU Power Cap: 470W (Combined 3 GPUs)

1. Hardware Summary

System component specification for origami (v2.0 server platform). Upgraded from the legacy X299 platform to a full enterprise AMD EPYC server architecture designed for massive multi-GPU Folding@home throughput and large local LLM inference.

Component Hardware Specification System Allocation & Details
CPU AMD EPYC 7543 (32 Cores / 64 Threads @ 2.80 GHz Base / 3.70 GHz Boost, 256MB L3 Cache) Socket SP3 Enterprise Server Processor (225W TDP, 128 PCIe 4.0 CPU Lanes)
Motherboard ASRock Rack ROMED8-2T (Rev 1.03, BIOS P3.80) Enterprise Server Board featuring 7x PCIe 4.0 x16 full-length slots, dual M.2 NVMe, dual OCuLink x4, and integrated ASPEED AST2500 BMC
RAM / Memory 256GB DDR4 Eight-Channel Registered ECC Memory (8x 32GB Samsung DDR4-3200 RDIMMs @ 2400 MT/s) Full 8-channel memory bandwidth saturation for intensive multi-GPU folding feeds and massive local LLM context windows
Power Supply (PSU) Corsair AX1600i 1600W 80+ Titanium Certified Dedicated 1600W continuous digital power delivery (80+ Titanium, GaN transistors) with full multi-GPU headroom
Chassis / Case Phanteks Enthoo Pro 2 Server Edition TG High-airflow full-tower chassis optimized for multi-GPU SSI-EEB / CEB / ATX server motherboards with side/front radiator and intake configurations
Storage (Primary OS) Samsung SSD 850 EVO 500GB (2.5" SATA 6Gb/s SSD) Retained from v1.5 platform for host OS, boot partition, and system services (/, /boot, /boot/efi)
Storage (Data / App) Samsung SSD 970 EVO Plus 2TB (M.2 PCIe 3.0 x4 NVMe SSD) Retained from v1.5 platform, mounted at /opt for Docker containers, local LLM models, and telemetry datasets
Networking Dual 10GBase-T RJ45 (Intel X550-AT2) + Dedicated 1Gbps IPMI Management LAN (Realtek RTL8211E) High-speed 10Gbps backbone link to local habitat network plus isolated BMC remote management
Cooling & Thermal Control High-Static Pressure PWM Thermal Control + Automated Dynamic Fan Regulation (gpufanpin) Continuous fan speed management pinning GPU blowers/fans to optimal duty cycles during live F@H execution

2. Physical to Logical Mapping

Mapping physical motherboard PCIe slots on the ASRock Rack ROMED8-2T layout to PCI Bus Addresses, NVIDIA software indexes, physical silicon GPU UUIDs, and active compute roles on host origami.

Motherboard Slot GPU Model PCI Bus ID NVIDIA Index Physical GPU UUID Operating Role & Notes
PCIE7 (Gen4 x16) ZOTAC GAMING GeForce RTX 5070 Solid OC (12GB GDDR7) 81:00.0 1 dd40d876-3e63-1967-9ae5-8227782a482e Active (180W Cap, Silicon Winner @ ~90k PPD/W)
PCIE2 (Gen4 x16) PNY NVIDIA RTX PRO 4000 Blackwell Workstation (24GB GDDR7) 82:00.0 2 e1600107-6dc5-4e55-1efe-0bf6fa945286 Active (145W Cap, New Aug 16 Addition, ~112k PPD/W)
PCIE1 (Gen4 x16) PNY NVIDIA RTX PRO 4000 Blackwell Workstation (24GB GDDR7) 01:00.0 0 6f99ccc4-f9cb-c840-105b-f4aeaed1b3eb Active (145W Cap, ~108k PPD/W)

Note: The Asus PRIME SFF-Ready OC GeForce RTX 5070 (18de45ee..) from the v1.5 build has been retired from the active folding fleet and repurposed inside the refurbished origami v1 chassis as a giveaway workstation.


3. Hardware Upgrade Roadmap

Platform evolution chronology for node origami.