Rackscale Solution
Liquid-Cooled RackScale AI Built for Production Density
Liquid-Cooled AI Rack Solutions
From LiquidMax® RackScale 64 to GB300 NVL72 RackScale 72 and open-rack RackScale 128, AMAX delivers high-density AI rack solutions built for next-generation training, inference, and large-scale AI factory deployments. With direct liquid cooling, rack-level power design, high-speed networking, validation, and deployment support, AMAX brings GPU-dense infrastructure from design to production readiness.
LiquidMax® RackScale 64
LiquidMax® RackScale 64 is a fully liquid-cooled 42U AI rack designed to support next-generation training and inference workloads. Built for dense GPU deployments, it integrates compute, cooling, and power in a compact footprint. This solution enables energy-efficient operation with the thermal control needed for sustained performance.
Key Features
- Supports up to 64x NVIDIA® Blackwell GPUs
- 8x 4U liquid-cooled compute servers with NVIDIA NVLink™ interconnect
- Integrated direct liquid cooling with passive cold plates
- CDU with redundant PSU and dual hot-swap pumps
- Vertical cooling distribution manifold for even flow
- Turnkey, rack-ready solution for faster deployment
- Optional support for NVIDIA Quantum-2 InfiniBand or Spectrum™-4 Ethernet
LiquidMax® RackScale 64
High-Density Liquid-Cooled AI Rack Solution
System Overview
| Component | Details |
|---|---|
| GPU | 64× NVIDIA Blackwell GPUs (8× 4U compute nodes) |
| Cooling | Direct liquid cooling with CDU and vertical distribution manifold |
| Architecture | 42U rack with 8× 4U liquid-cooled servers |
| Power | Integrated power distribution with redundant pump and PSU design |
| Management | Out-of-band IPMI, in-band Ethernet switching, optional leak detection |
AceleMax® AXG-428IL – 4U Liquid-Cooled GPU Server
| Component | Details |
|---|---|
| CPU | Dual Intel Xeon 6700P/6500P/6700E series (LGA 4710) |
| Memory | 32× DDR5 DIMM slots (2DPC), up to 128GB per DIMM, 6400MT/s (1DPC) |
| GPU | NVIDIA Blackwell 8-GPU with NVIDIA NVSwitch |
| Storage | 8× NVMe (PCIe Gen 5) hot-swap bays, plus 2× from CPU, 1× internal M-key |
| Expansion | 8× HHHL + 2× FHHL PCIe Gen 5 x16 slots |
| Networking | 2× 1GbE RJ45 (Intel i350), 1× dedicated IPMI port |
Meeting the Power Demands of Modern AI
Today’s AI workloads require significantly more power than previous generations. As model sizes grow, so do the demands on power and thermal systems. The RackScale 64 is engineered to address these requirements with efficient liquid cooling that minimizes energy waste and maintains consistent performance during sustained high-power operation.
LiquidMax® RackScale 72 with NVIDIA GB300 NVL72
36 Grace™ Arm® Neoverse V2 CPUs and 72 Blackwell Ultra GPUs
LiquidMax® RackScale 72 brings NVIDIA GB300 NVL72 into a fully liquid-cooled rack architecture for high-density AI factories. With 72 Blackwell Ultra GPUs, 36 Grace CPUs, NVLink switching, 800Gb/s networking, and rack-level thermal design, AMAX delivers a production-ready platform for large-scale AI training and inference.
- Per Rack:
- 18 x Compute Tray + 9 x Switch Tray
- Fully Liquid-cooled
- Per Tray:
- 4 x NVIDIA Blackwell Ultra GPU + 2 x Grace™ CPU
- 960GB LPDDR5X + 1152GB HBM3e
- 4x NVIDIA NVLink™ Switch
- 2 x NVIDIA ConnectX®-8 NICs (1-port 800Gb/s)
- 1 x BlueField®-3 DPU (400Gb/s)
LiquidMax® RackScale 72 with NVIDIA Vera Rubin NVL72
36 Vera CPUs and 72 Rubin GPUs in a Rack-Scale System
NVIDIA Vera Rubin NVL72 is a liquid-cooled, rack-scale AI computing platform combining 72 Rubin GPUs and 36 Vera CPUs through high-bandwidth NVLink. Designed for large-scale training, reasoning, agentic AI, and inference, the platform operates as a unified rack-level compute system and can scale across multiple racks to form production AI factories.
Features
- Powered by NVIDIA NVLink™ 6 Switch, NVIDIA ConnectX®-9 SuperNIC, and NVIDIA BlueField®-4 DPU
- Cable-less, fully liquid-cooled design for high efficiency
- Modular architecture with SOCAMM memory for easy serviceability
LiquidMax® RackScale 128
The LiquidMax® RackScale 128 is a fully liquid-cooled 51OU AI rack built with open rack architecture. Designed for hyperscale deployments, it delivers the GPU density, thermal efficiency, and power infrastructure required to support large-scale AI and ML workloads.
Key Features
- Supports up to 128x NVIDIA Blackwell GPUs or AMD Instinct™ MI355X / MI325X accelerators
- 8x 5OU compute trays with cold plate liquid cooling
- Centralized 48V busbar power distribution for improved efficiency
- Built-in leakage detection, CDU support, and optional RDHx integration
- Supports dual 400W CPUs and high-performance DDR5 memory
- PCIe Gen 5 expansion for high-speed networking and connectivity
LiquidMax® RackScale 128
High-Density Liquid-Cooled AI Rack Solution
System Overview
| Component | Details |
|---|---|
| GPU | 128× GPUs (8× 5OU compute trays) |
| Cooling | Liquid-to-liquid cooling with CDU support and active leakage detection |
| Architecture | Open rack design, 51OU rack with 8× 5OU trays |
| Power | 8× 33kW power shelves with centralized 48V busbar distribution |
| Management | Integrated leakage detection trays and monitoring |
Compute Node Specifications
| Component | Details |
|---|---|
| CPU | 2× AMD Turin 9655, 96-core, SP5 (2-socket), up to 400W TDP |
| Memory | 24× DDR5 DIMM slots, 96GB RDIMM (5600MT/s), 1DPC |
| Expansion Slots |
6× PCIe 5.0 x16 (SS FHHL):
|
Built for Hyperscale AI
The RackScale 128 is designed for large-scale cloud providers, data centers, and enterprises with extreme compute requirements. The open rack architecture design supports wider 21″ racks for improved airflow and accessibility, while front-to-back cooling and blind-mate manifolds simplify installation and maintenance.
Its cold plate liquid cooling ensures stable operation under sustained load. The 48V power busbar architecture improves electrical efficiency, minimizes cabling, and enhances serviceability.
21″ Design Advantages
- Modular 5OU tray design for scalable deployment
- Front-access compute and power nodes
- Shared busbar reduces complexity
- Blind-mate liquid manifolds for efficient maintenance
- 21” chassis format increases airflow compared to traditional 19” racks
OCP Open Rack V3 for High-Density AI
AMAX supports OCP Open Rack V3 architecture for high-density AI infrastructure, combining 21-inch open rack design, 48V DC power distribution, liquid cooling, and rack-level system engineering for scalable GPU deployments.
Open Architecture for Rack-Scale AI
21-Inch Open Rack Architecture
The wider Open Rack design provides additional space for high-density compute, accelerator, power, and liquid-cooling hardware.
48V DC Busbar Power
Rack-level 48V DC distribution supports direct busbar connectivity and reduces dependence on individual rack power cabling.
Blind-Mate Power Connectivity
ORV3 supports direct power connection to the rack busbar as compatible equipment is inserted, reducing manual cabling and improving serviceability.
Liquid Cooling Ready
ORV3 supports rack-level liquid cooling with manifolds, cold plates, and CDU or facility-side coolant connections for high-density AI systems.
Designed for High-Density Compute
The modular rack architecture supports GPU, accelerator, storage, and other high-power infrastructure used in AI and HPC environments.
AMAX Rack-Level Engineering
AMAX combines engineering expertise with customization capabilities to transform OCP ORV3 architecture into customer-specific rack-scale solutions, from initial design and validation to manufacturing and production deployment.