AVAILABLE AND IN STOCK
Accelerate On-Prem AI with AMAX
High-density, rack-scale solution engineered for large-scale enterprise AI workloads with 32 NVIDIA HGX™ B300 GPUs
Accelerate On-Prem AI with AMAX
High-density, rack-scale solution engineered for large-scale enterprise AI workloads with 32 NVIDIA HGX™ B300 GPUsg
Datasheet Request a QuoteKey Features
Tensor/Transformer Cores
Up to 576 specialized cores that accelerate transformer operations (e.g., attention layers) critical to LLM performance.
High Memory Capacity
Up to 8.4TB of HBM3e memory, enabling efficient handling of massive models and long context windows without bottlenecks.
In-Node Bandwidth
NVIDIA NVLink/NVSwitch interconnect keeps GPUs fully utilized across large-scale training workloads.
High-Speed Networking
800Gbps NVIDIA InfiniBand for rapid multi-node synchronization across distributed AI clusters.
Networking
HBM3e GPU memory
Fully Configured for AI at Scale
The AMAX RackScale 32 with NVIDIA HGX B300 is a complete rack solution engineered, assembled, and validated by AMAX. Our teams manage the full process from design to manufacturing to deployment, delivering a solution that arrives complete with compute, networking, and storage ready for activation.
Each rack includes four 8U compute servers accelerated by 32 Blackwell Ultra GPUs, supported by 800Gbps InfiniBand networking and a high performance storage platform. The solution follows the NVIDIA reference design and is brought together through AMAX engineering, quality assurance, and deployment services to help customers move directly into large scale AI development.
Engineering Expertise
Solution Architects
AMAX's solution architects optimize IT configurations for performance, scalability, and industry-specific reliability.
Networking
AMAX designs custom networking topologies to enhance performance and security in AI and HPC environments.
Thermal Management
AMAX implements innovative cooling technologies that boost performance and efficiency in dense computing setups.
Compute Optimization
AMAX ensures maximum performance through benchmarking and testing, aligning hardware and software for AI workloads.
AceleMax® AXG-828U
The AceleMax® AXG-828U pairs 8x NVIDIA HGX B300 GPUs with dual Intel Xeon 6700-series CPUs, PCIe Gen 5 NVMe storage, and 12 expansion slots. Designed for dense AI deployments, it delivers high-speed interconnects and efficient thermal control in a 8U rackmount form factor.
Datasheet
AI Training & Inference Performance
Next Level Training Performance
The second generation Transformer Engine with FP8 enables 4x faster training for large models like Llama 3.1 405B. Combined with NVLink at 1.8 TB/s, InfiniBand networking, and Magnum IO software, it scales efficiently across enterprise clusters.
Real Time Inference
HGX B300 delivers up to 11x higher inference performance over the Hopper generation. Blackwell Tensor Cores with TensorRT LLM innovations accelerate inference for Llama 3.1 405B and other large models.
Fully Managed AI Deployment
AMAX’s approach to AI solutions begins with intelligent design, emphasizing the creation of high-performance computing and network infrastructures tailored to AI applications. We guide each project from concept to deployment, ensuring systems are optimized for both efficiency and future scalability.
Activate Your AI Infrastructure Instantly with HostMax™
HostMax™ is AMAX’s in-house deployment service that lets you power on and operate your liquid-cooled AI systems as soon as they’re built. Instead of waiting for colocation space, HostMax™ provides immediate hosting at AMAX’s facility, enabling a direct transition from assembly to deployment for testing, validation, and early production.
Learn More
AMAX RackScale 32
AMAX RackScale 32 combines compute performance with efficient scalability for industries including healthcare, finance, public sectors, manufacturing, neoclouds, and research institutions.
- Dual Socket Intel® Xeon® 6700E/6700P series processors
- 32x NVIDIA Blackwell Ultra GPUs
- Up to 8.4TB total HBM3e GPU memory per rack
- Total FP4 Tensor Core 576 PFLOPS