NVIDIA GB200 NVL72 Liquid-to-Air Rack Solutions

Overview

The LiquidMax® ALC-B4872 GB200 NVL72 AI POD is pre-configured with 36 Grace Blackwell Superchips featuring 72 NVIDIA Blackwell GPUs and 36 Grace CPUs, interconnected by 5th-generation NVLink. This advanced solution is designed to deliver cutting-edge performance for workloads such as:

Large Language Model (LLM) Inference
Retrieval-Augmented Generation (RAG)
High-speed data processing

With its scale-out, single-node NVIDIA MGX architecture, AMAX’s LiquidMax® ALC-B4872 GB200 NVL72 AI POD enables a wide variety of system designs and networking options to integrate into existing data center infrastructure.

In addition, the LiquidMax® ALC-B4872 GB200 NVL72 AI POD includes NVIDIA BlueField®-3 data processing units to enable cloud network acceleration, composable storage, zero-trust security and GPU compute elasticity in hyperscale AI clouds. The LiquidMax® ALC-B4872 GB200 NVL72 AI POD provides up to a 30x performance increase compared to the same number of NVIDIA H100 Tensor Core GPUs for LLM inference workloads, and reduces cost and energy consumption by up to 25x.

💡

POD Configuration

Key Features

Up to 30x Performance Increase for LLM inference workloads
25x Reduction in Cost and Energy Consumption compared to NVIDIA H100
Scale-Out, Single-Node NVIDIA MGX Architecture for flexible deployment options
Efficient Liquid-to-Air (L2A) Cooling for high-density AI and HPC workloads

Per Rack

Component	Details
Compute Trays	18
Switch Trays	9
Cooling	Liquid-to-Air (L2A)

Per Tray

Component	Details
NVIDIA GPUs	4 NVIDIA Blackwell GPUs
CPUs	2 Grace CPUs
Memory	960GB LPDDR5X + 288GB HBM3e
Networking	2x NVIDIA ConnectX®-7 NIC (400GbE) + 2x BlueField®-3 DPU NICs (400GbE)

Grace Blackwell Superchip Performance

30X

LLM Inference

vs. NVIDIA H100 Tensor Core GPU

LLM Training

vs. H100

25X

Energy Efficiency

vs. H100

18X

Data Processing

vs. CPU

Blackwell GPU Performance

Metric	Per Rack	Per Tray
FP4 Tensor Core	1,440 PFLOPS	40 PFLOPS
FP8/FP6 Tensor Core	720 PFLOPS	20 PFLOPS
INT8 Tensor Core	720 POPS	20 POPS
FP16/BF16 Tensor Core	360 PFLOPS	10 PFLOPS
TF32 Tensor Core	180 PFLOPS	5 PFLOPS
FP32	6,480 TFLOPS	180 TFLOPS
FP64	3,240 TFLOPS	90 TFLOPS

Grace CPU Specifications

Metric	Per Rack	Per Tray
CPU Core Count	2,592 Arm® Neoverse V2 cores	72 cores
CPU Memory	Up to 17 TB LPDDR5X	Up to 480GB LPDDR5X
Memory Bandwidth	Up to 18.4 TB/s	Up to 512 GB/s

Per Rack Memory and Bandwidth

Up to 13.5 TB HBM3e with bandwidth up to 576 TB/s
NVLink Bandwidth: Up to 130 TB/s

Liquid-to-Air Cooling Solution

AMAX’s LiquidMax® ALC-B4872 GB200 NVL72 AI POD with Liquid-to-Air (L2A) Cooling provides an efficient thermal management solution for high-performance data centers.

The system circulates liquid coolant to absorb heat from high performance components within the rack. This heat is then transferred to the air via a sidecar cooling unit, which expels the hot air into the Computer Room Air Conditioner (CRAC) for final dissipation. This efficient design provides a cost-effective and scalable cooling solution for modern data centers.

Liquid-to-Air CDU

The LiquidMax® ALC-B4872 GB200 NVL72 AI POD uses an innovative Liquid-to-Air (L2A) cooling solution:

Cooling Capacity: Up to 240 kW for 2-4 racks
No need for facility liquid integration
Scalable and cost-effective for modern data centers

Specification	CDU Option 1	CDU Option 2
Power Consumption	11.32 kVA	11.32 kVA
Racks Managed	Up to 2	Up to 4

Why Choose AMAX?

AMAX combines innovative engineering with industry-leading hardware such as the NVIDIA GB200 NVL72 to deliver scalable solutions tailored to your unique needs. From design to deployment, AMAX provides expert support every step of the way.

10+ Years Experience with HPC Liquid Cooling
On-Site Installation & Cluster Bring-up
Network Topology
Bi-directional Logistics
Testing & Validation
Maintenance & Upgrades
Troubleshooting & Repair Services

💡