NVCL-manager-v9.42.113

GPU Nodes & Cluster Pool

Showing active SuperPOD racks & high-density HGX nodes (Total: 4,000 B200 & 1,200 H200)

Node Name	GPU Model & Qty	Status	Util (Load)	Memory (VRAM)	Power & Temp	Active Workload	Actions

Cluster Inventory: 4,000x NVIDIA B200 + 1,200x NVIDIA H200 = 5,200 Units

Cluster Thermal & Power

Live 60Hz

B200 Core (69°C) H200 SXM (64°C) Coolant (32°C)

Memory Footprint (937.2 TB)

89.4% Allocated

B200-P1

B200-P2

H200-P1

H200-P2

NVLink Bandwidth

Per-GPU Fabric

1.8 TB/s NVLink 5.0 Bi-Dir

BLACKWELL ARCHITECTURE

NVIDIA B200 SXM6 Pool — 4,000 Units

4,000 GPUs

500 Racks × 8 GPUs

Total HBM3e

768.0 TB

192GB / GPU

FP4 Tensor Peak

80.0 ExaFLOPS

20 PFLOPS / GPU

Power Budget

4.00 MW

1000W TDP / GPU

MIG / Multi-Instance Slicing Mode Full Instance (8x192GB Unified)

HOPPER ARCHITECTURE

NVIDIA H200 SXM5 Pool — 1,200 Units

1,200 GPUs

150 Racks × 8 GPUs

Total HBM3e

169.2 TB

141GB / GPU

Memory Bandwidth

4.8 TB/s

Per GPU SXM5

Power Budget

840 kW

700W TDP / GPU

MIG / Multi-Instance Slicing Mode Inference Optimized (7x20GB MIG)

Slurm / Kubernetes Distributed Job Queue

Dispatching workloads across 4,000 NVIDIA B200 and 1,200 NVIDIA H200 GPUs

Job ID	Workload Name	Owner	Allocated GPUs	Progress	State	Controls

Production Inference Endpoints (TensorRT-LLM / vLLM)

Low-latency model serving backed by H200 & B200 NVLink clusters

Real-Time Kernel & DCGM Cluster Audit Stream

Live hardware events from NVCL-manager-v9.42.113

Operator & Security Controls

System Build:NVCL-manager-v9.42.113

Authenticated User:cto-prista

Session Guard:SHA-256 Token Active

Anti-Inspect Deterrent:Enabled

Cluster Thermal & Power Protection

Auto-Throttle GPU Clocks at 82°C Redundant NVLink Failover Routing Direct-to-Chip Liquid Cooling Boost (100% Pump PWM)