Explore accelerators, storage and memory, then follow the connections from a single server to a complete cluster.
Estimate GPUs, memory, storage, power and budget with adjustable assumptions.
Go deeper into the components behind the architecture. Each guide includes product examples and links to manufacturer documentation.
Selected manufacturer specifications. Product announcements do not establish local stock, delivery dates or Ardelvo reseller status.
B300 is the accelerator; DGX B300 is an eight-GPU server. GB300 NVL72 combines Blackwell Ultra GPUs and Grace CPUs across a liquid-cooled rack.
Planning implication A server and an NVL72 rack require different power, cooling and service arrangements. Specify the complete system, not just the GPU name.
NVIDIA DGX specifications GB300 rack design
A high-memory accelerator for AI and HPC, with support for microscaling data types. Evaluate the ROCm software path alongside the hardware configuration.
Planning implication Validate your framework, kernels and model-serving stack on the intended software release before committing to a platform.
AMD product specificationsNew designs bring the CPU, accelerator, networking and cooling closer together. Procurement and facility planning increasingly need to happen at the same time.
These simplified logical diagrams explain system relationships. They are planning illustrations, not vendor wiring diagrams or construction designs.
The model, precision and context length affect memory demand. GPU-to-GPU links affect how efficiently work can be divided inside a server.
HBM is high-bandwidth memory attached to the accelerator. Aggregate capacity does not automatically behave like one shared memory pool.
When a workload spans nodes, communication and storage can limit useful GPU throughput. Design the fabric around traffic patterns and failure tolerance.
Node counts and links are illustrative. Select topology, optics and adapters against the chosen vendor design.
Rack-scale systems join many accelerators through a scale-up fabric. The facility must support the complete delivered configuration, including service access and cooling interfaces.
A CDU is a coolant distribution unit. Exact cooling and power requirements are platform-specific.
Further reading: NVIDIA NVL72 reference architecture · DGX B300 system guide
Compare systems against a representative workload. Peak FLOPS alone cannot tell you what a production deployment will deliver.
Account for model weights, activations, KV cache and runtime overhead. Training and inference can have very different requirements.
Check supported frameworks, drivers, kernels and serving tools. Run your model on the intended software versions.
Separate GPU memory bandwidth, in-system interconnect and network bandwidth. Each addresses a different bottleneck.
Include facility changes, installation, spares, support and energy in the evaluation—not only the equipment price.
Share the workload, preferred platforms and site constraints.
What are you working on? I can point you to the right information and help prepare your inquiry.
Guided answers from Ardelvo’s service information. No live pricing or stock. This conversation stays in this page. Draft transfer briefly uses browser session storage; you review before sending. Do not enter confidential information.