| # TECHNICAL SPECIFICATION: JiRack 236B Frontier Cluster | |
| **Document ID:** CMS-JR-236B-PROC-2025 | |
| **Project:** JiRack Balanced Frontier Deployment | |
| **Lead Architect:** Konstantin Vladimirovich Grabko | |
| --- | |
| ## 1. COMPUTE SYSTEM (High-Density GPU Nodes) | |
| The deployment requires a minimum of two (2) fully populated 8-GPU nodes to achieve the necessary VRAM and throughput for the 108-layer dense architecture. | |
| | **Component** | **Minimum Specification** | **Required Feature for JiRack** | | |
| |--------------------|-----------------------------------|---------------------------------------------------------| | |
| | **Chassis** | 6U - 8U Rackmount (e.g., NVIDIA HGX H100) | Support for 700W SXM5 GPUs | | |
| | **Accelerator** | 16x NVIDIA H100 80GB (SXM5) | 1.28 TB aggregate HBM3 VRAM | | |
| | **GPU Interconnect** | NVLink + NVSwitch (900GB/s) | High-width Tensor Parallelism (TP) | | |
| | **Processor (CPU)** | 2x Intel Xeon Platinum 8480+ (per node) | 224 total cores for BRE routing | | |
| | **System Memory** | 2TB DDR5-4800MHz ECC | Weight-caching for 108 layers | | |
| | **Local Storage** | 8x 3.84TB NVMe U.2 (RAID 0) | Sequential read for fast model loading | | |
| --- | |
| ## 2. NETWORKING & FABRIC (Low-Latency Backplane) | |
| Because the 236B model requires frequent cross-node communication (Pipeline Parallelism), **standard Ethernet is unacceptable**. | |
| - **Primary Fabric:** NVIDIA Quantum-2 InfiniBand (NDR 400Gb/s). | |
| - **Adapters:** 8x ConnectX-7 NDR InfiniBand cards per node (1:1 GPU-to-NIC ratio). | |
| - **Switching:** 1x QM9700 Quantum-2 Switch (64 ports of NDR). | |
| - **Features:** Must support GPUDirect RDMA and SHARPv3 (Scalable Hierarchical Aggregation and Reduction Protocol) to offload 14:1 GQA reduction tasks. | |
| --- | |
| ## 3. POWER & THERMAL MANAGEMENT | |
| Operating a 236B Frontier model at 100% duty cycle generates extreme thermal load. | |
| | **Parameter** | **Specification** | | |
| |-----------------------|---------------------------------------------------------| | |
| | **Peak Power Draw** | 10.2 kW per server (~22kW total including switching) | | |
| | **Voltage** | 200V - 240V AC (3-Phase Delta/Wye) | | |
| | **Cooling Type** | Rear Door Heat Exchanger (RDHx) or Liquid Cooling (D2C) | | |
| | **Ambient Ops** | 18°C – 24°C (64°F – 75°F) with N+1 Redundancy | | |
| | **Rack Rating** | 48U Heavy-Duty (Static Load: 1500kg+) | | |
| --- | |
| ## 4. SOFTWARE STACK & COMPLIANCE | |
| - **OS:** Ubuntu 22.04 LTS (NVIDIA-Certified). | |
| - **Driver:** NVIDIA Data Center Driver (Current 5xx.x branch). | |
| - **Middleware:** CUDA 12.x, NCCL 2.x, and cuDNN. | |
| - **Security:** TPM 2.0 and Secure Boot (Mandatory for BRE architecture protection). | |
| --- | |
| ## 5. LICENSING & ATTRIBUTION (MANDATORY) | |
| - **Vendors and Integrators:** Notified that this hardware is intended to host the JiRack 236B Architecture. | |
| - **IP Ownership:** All architectural logic remains the property of Konstantin Vladimirovich Grabko. | |
| - **Compliance:** The system must allow the `proof_of_authorship` buffer to be verified upon request. | |
| - **Commercial Terms:** Deployment is subject to the Commercial License Agreement V.1.2 (**5% Net Revenue Royalty**). |