11 KiB
11 KiB
Master Engineering Specification & Architecture Baseline: Open-Source 6U 19-Inch Tri-Mode Storage Array
Document Version: 1.0
Target Environment: Open-Source Hardware / KiCad 8 / EIA-310-D 19-Inch 6U Rackmount
Author: Hardware Engineering Architecture Team
1. Executive Summary & Design Philosophy
This document serves as the authoritative baseline specification for an open-source, enterprise-capable 6U 19-inch rackmount hybrid storage array. The architecture is engineered to challenge expensive, proprietary Storage Area Networks (SANs) by pairing democratized compute silicon, enterprise-grade storage controllers, and low-cost manufacturing methods.
Core Strategic Targets:
- Bill of Materials (BOM) Target: Carrier Baseboard + Compute Engine < $100 (est. $92.50); Mechanical Skeleton & Facade < $65 (est. $61.50). Total core unit target: ~$155–$175.
- Universal Storage Protocol Compatibility: Native support for SAS (12Gb/s), SATA (6Gb/s), and PCIe NVMe (U.2/U.3 SFF-TA-1001) within the same physical bays without converter adapters.
- Manufacturing Democratization: 4-layer standard FR4 PCB substrate, 2020 T-slot aluminum extrusion internal frame, and modular interlocking 3D-printed facade parts sized for standard desktop 3D printers (
\le 220\,\text{mm} \times 220\,\text{mm}build plates). - Rail-Less Mounting: Uses heavy-gauge steel rack ears and static 4-post shelves/L-runners (or OCP OrV3 sled mounting) to eliminate expensive, finicky telescoping server rails.
2. System Architecture & Silicon Selection
+-----------------------------------------------------------------------------------+
| 6U CHASSIS ENVELOPE |
| |
| +------------------------+ +---------------------+ +----------------+ |
| | Radxa CM3 (RK3568) | | Broadcom Tri-Mode | | 8x U.3 Drives | |
| | - Quad Cortex-A55 | PCIe | SAS3408 / SAS3416 | MCIO | (NVMe/SAS/SATA)| |
| | - 1 TOPS NPU | Gen3 | - Dual Cortex-A15 |=====>| - Hot-Swap Sled| |
| | - 4GB/8GB LPDDR4 | x2 | - Tri-Mode SerDes | Twin | - SFF-TA-1001 | |
| | - Dual GbE NICs |=====>| - DataBolt Aggreg. | Ax | Universal Bay| |
| +-----------+------------+ +----------+----------+ +----------------+ |
| | UART/I2C | Sideband/SGPIO |
| v v |
| +-----------------------------------------------------+ |
| | Raspberry Pi RP2040 Auxiliary Management MCU | |
| | - Staggered Spin-Up (PUIS) Power Sequencing | |
| | - 4-Pin PWM Closed-Loop Thermal PID Fan Control | |
| | - Multi-Point I2C Environmental Telemetry | |
| +-----------------------------------------------------+ |
+-----------------------------------------------------------------------------------+
2.1 Compute Subsystem: Radxa Compute Module 3 (RK3568)
- SoC: Rockchip RK3568 (Quad-core ARM Cortex-A55 @ 2.0 GHz, 64-bit).
- AI Acceleration: Integrated 1.0 TOPS NPU utilized for on-device S.M.A.R.T. telemetry analysis, predictive drive degradation heuristics, and automated caching tier algorithms.
- Memory: 4GB to 8GB LPDDR4 @ 3200 MT/s (sufficient for Linux OS and ZFS ARC metadata caching).
- I/O Topology:
- 1x PCIe Gen 3.0 (x2 lanes, 16 Gbps raw / ~1.97 GB/s theoretical) dedicated as the upstream host link to the storage controller.
- 1x PCIe Gen 2.0 (1 lane) available for auxiliary 2.5GbE/10GbE network expansion.
- Dual GbE interfaces directly exposed.
- Idle Power:
\le 2.0\,\text{W}system idle.
2.2 Storage Controller Silicon: Broadcom SAS3408 / SAS3416
- Architecture: 6th Generation Tri-Mode I/O Controller (IOC) featuring embedded dual-core ARM Cortex-A15 @ 1.2 GHz.
- Performance: >1,000,000 IOPS; manages up to 2,000 devices via Fusion-MPT architecture without requiring dedicated controller DDR RAM.
- Tri-Mode SerDes: Hardware auto-negotiation per PHY between SAS 12G/6G/3G, SATA 6G/3G, and PCIe NVMe 8GT/s (Gen 3).
- DataBolt™ Technology: Buffers and aggregates 6Gb/s SATA/SAS data streams into full 12Gb/s fabric bandwidth to prevent slow disks from bottlenecking high-speed channels.
2.3 Auxiliary Management Microcontroller: Raspberry Pi RP2040
- Role: OpenBMC-lite physical layer supervisor running deterministic C/C++ firmware.
- Responsibilities:
- Power-Up In Standby (PUIS) & Staggered Spin-up: Sequences drive spin-up via SAS/SATA Pin 11 to avoid inrush current collapse on the 12V power rail.
- Closed-Loop Thermal Management: Reads multi-point I2C temperature sensors (TMP117 / MCP9808) placed across the drive bay, controller heatsink, and power stage; drives 4-pin PWM enterprise fan curves.
- Host Interface: Reports environmental metrics, fan RPMs, and rail voltages to the RK3568 host via UART/I2C.
3. High-Speed Interconnect & PCB Strategy
3.1 4-Layer FR4 Baseboard Design (KiCad 8)
- Target Substrate: Standard 4-layer FR4 (e.g., JLC04161H-7628 stack-up, 1.6mm thickness, 1 oz copper weight).
- Stack-Up: L1 (High-Speed Signal / Components) - L2 (Solid GND Reference) - L3 (12V/5V/3.3V Power Planes) - L4 (Low-Speed Signal / Auxiliary Power).
- Controlled Impedance Targets:
- PCIe Gen 3 Differential Pairs:
85\,\Omega \pm 10\% - SAS 12G Differential Pairs:
100\,\Omega \pm 10\%
- PCIe Gen 3 Differential Pairs:
- Trace Length Mitigation: High-speed traces on the FR4 board are strictly constrained to
\le 4\,\text{inches}(100 mm) between chip pins and connector pads. - Cable Offloading Architecture: High-speed SerDes lines are broken out immediately into high-density SFF-TA-1016 (MCIO) or SlimSAS surface-mount connectors. Inexpensive twinaxial copper ribbon cables transport signals directly to the backplane, eliminating the need for expensive 8–10 layer low-loss PCB substrates.
4. Power Delivery Network (PDN) & PLP Architecture
4.1 Native 12V DC Decoupled Bus
- Main Power Bus: 12V DC unified main rail fed by external mass-produced commodity power bricks.
- Input Interfaces:
- Profile A (Drive-Only / Low-Wattage): 12V 10A–15A (120W–180W) standard DC barrel jack.
- Profile B (GPU Growth / High-Wattage): High-current Molex Mini-Fit Jr. 6-pin/8-pin receptacle (handles up to 300W–500W external supply).
- Front-End Protection: Ideal Diode Controller IC (e.g., TI LM74700-Q1 or LTC4357) on the 12V input to protect against reverse polarity, overvoltage spikes, and back-feeding.
- Local Step-Down Regulation: High-efficiency synchronous buck converters on the carrier PCB:
12\,\text{V} \rightarrow 5.0\,\text{V} \ @ \ 10\,\text{A}(HDD logic boards & USB subsystem).12\,\text{V} \rightarrow 3.3\,\text{V} \ @ \ 5\,\text{A}(Radxa CM3, Broadcom IOC logic, RP2040, backplane sideband).
4.2 Power Loss Protection (PLP) Circuit (Open Engineering Target)
- Goal: Hardware supercapacitor (or tantalum polymer) energy storage bank that provides sufficient holdup time during sudden AC mains loss.
- Target Function: Keep the 12V/5V/3.3V rails alive for
\sim 50\,\text{ms} - 250\,\text{ms}while asserting an emergency interrupt to the RK3568 and Broadcom IOC to flush in-flight volatile write caches to non-volatile media.
5. Mechanical Enclosure & Thermal Architecture
+-----------------------------------------------------------------------------------+
| FRONT (Intake) REAR (Exhaust)|
| |
| [Modular 3D Facade] -> [8x 3.5"/U.3 Bays] -> [Fan Wall] -> [Open Mezzanine Area] |
| (PETG/ASA Dovetails) (Front 4-5 inches) (3x 120mm) (10-14" GPU/Compute) |
+-----------------------------------------------------------------------------------+
5.1 Hybrid Structural Skeleton
- Primary Frame: 2020 (20mm x 20mm) T-slot / V-slot anodized aluminum extrusions.
- Outer Skin: Flat 1.0mm–1.5mm galvanized steel or aluminum sheet metal panels (simple laser/waterjet cut rectangles, no complex stamping).
- Fastening System: M3 and M4 drop-in extrusion T-nuts paired with heat-set brass threaded inserts in all plastic components (zero direct tapping into polymer).
- Rack Mounting: Heavy-gauge 3mm steel 6U rack ears bolted directly into frame corners; slides onto static 4-post heavy-duty shelves or L-runners (bypassing telescoping rails).
5.2 3D-Printable Facade & Modularity
- Material: PETG, ABS, or ASA (PLA strictly forbidden due to thermal creep under 50°C+ ambient server loads).
- Print Bed Optimization: Front facade split into 3–4 interlocking sections utilizing dovetail joints, allowing fabrication on standard 220mm desktop 3D printers.
- Internal Expansion: Front 4–5 inches house the drive cage and carrier backplane; rear 10–14 inches provide an unobstructed vertical mezzanine compute area for full-height PCIe/GPU accelerator cards.
6. Open Technical Questions & Directives for Deep Research
The following subsystems require exhaustive engineering research, mathematical validation, and component selection:
- PLP Circuitry & Energy Calculations:
- Exact mathematical derivation of required capacitance
C = \frac{2 \cdot P \cdot t}{\eta \cdot (V_{\text{start}}^2 - V_{\text{drop}}^2)}for a 120W peak load with a 100ms holdup window. - Component selection (LTC3350, LTC4041, or TPS25982 eFuse) with exact Manufacturer Part Numbers (MPNs) and sub-$8 budget target.
- Exact mathematical derivation of required capacitance
- High-Speed Pinout & Routing Specification:
- Complete pin-to-pin schematic mapping from Radxa CM3 (PCIe 3.0 x2) to Broadcom SAS3408/3416 host pins.
- SFF-TA-1016 (MCIO) to U.3 (SFF-TA-1001) backplane SerDes breakout wiring table.
- High-speed signal integrity layout rules in KiCad 8 (via stitching, antipad sizing, AC coupling cap placement).
- RP2040 OpenBMC Firmware Architecture:
- State machine design for PUIS staggered spin-up delays (e.g., 2-second staggered spin interval per drive).
- Closed-loop PID fan curve algorithm based on multi-zone I2C temperature inputs.
- Host communication protocol specification (UART/I2C command structure).
- Thermal CFD & Airflow Budgeting:
- Total CFM and static pressure (
\text{mm}\,\text{H}_2\text{O}) required to pull air through dense 8-drive mechanical/NVMe arrays and cool high-power components. - Selection of 80mm / 120mm enterprise PWM fans with optimal acoustic and static pressure profiles.
- Total CFM and static pressure (