Dell Pro Max with GB300

Key features of the Dell Pro Max with GB300 workstation

GB300 Grace Blackwell Ultra superchip and 748GB of coherent memory

At the heart of the Dell Pro Max with GB300 is the NVIDIA GB300 Grace Blackwell Ultra superchip – a unique combination of a Blackwell Ultra GPU with 252GB of HBM3e memory and 7.1TB/s throughput, and a 72-core Grace Neoverse V2 ARM processor with 496GB of LPDDR5X and 396GB/s throughput. Both parts of the superchip are connected by NVLink-C2C technology with 900GB/s bandwidth, which eliminates the traditional bottleneck between CPU and GPU memory and creates a unified 748GB coherent memory space. The result is 20 petaFLOPS of performance in FP4 format and 330 TOPS in Int8 format – precisely where an entire server room floor was previously required.

NVIDIA MIG and ConnectX-8 for AI team performance in a single device

The Dell Pro Max with GB300 is not just a personal supercomputer, but a powerful local compute node for the entire AI team. NVIDIA MIG technology allows you to partition the GPU into up to seven fully isolated instances, each with its own dedicated memory, compute cores, and network connection, for multiple researchers to work concurrently without impacting each other. Connecting two stations via an NVIDIA ConnectX-8 SuperNIC with up to 800Gb/s bandwidth then doubles the performance to over 40 petaFLOPS and over 1.4TB of coherent memory to work with the largest available models – without the need to build an expensive server cluster.

AI application prototyping and development without limits

The complete NVIDIA AI software stack, combined with NVIDIA OpenShell, provides developers with a full-featured platform for creating AI models, AI agents, and AI-augmented applications – all locally, without cloud dependency. The NVIDIA DGX OS comes pre-configured with CUDA-X libraries, Container Toolkit, Docker, and NVIDIA AI Enterprise for an immediate start to development. Once the solution is ready for production deployment, the Dell Pro Max with GB300 enables direct and seamless migration of workloads to the NVIDIA DGX Cloud or other NVIDIA-accelerated infrastructure – with the same tools, libraries, and pre-trained models.

Fine-tune AI models with billions of parameters right at your workstation

Use the Dell Pro Max with GB300 to fine-tune pre-trained models with billions of parameters right at your workstation. Training on your own data allows you to specialise AI models for specific industry data or use cases – without sending sensitive data to the cloud or paying for expensive GPU instances. The NemoClaw platform then ensures the secure and isolated operation of autonomous AI agents with full governance control. The result is a model tailored exactly to your scenario, created locally, securely, and under your complete control.

Dell Pro Max with GB300 provides data science at the highest level

The combination of 748GB of coherent memory and 20 petaFLOPS of performance makes the Dell Pro Max with GB300 the ideal workstation for demanding data analytics and machine learning projects. Extensive data sets, complex computational models, and demanding training processes that previously required a powerful cloud cluster or a dedicated server now run directly on your desk. Four M.2 NVMe slots – two PCIe 5.0 x2 and two PCIe 6.0 x4 for training data – two 400Gb/s QSFP network interfaces and three PCIe 5.0 x16 slots ensure data throughput worthy of a research infrastructure – fast, local, and without waiting for remote infrastructure.

Inference of models with up to 1 trillion parameters

The Blackwell Ultra architecture's Tensor Cores, with FP4 format support, achieve up to 20 petaFLOPS of performance. Combined with 748GB of coherent memory, this lets you run inference for the largest available AI models with up to 1 trillion parameters right at your desk. Test, validate and deploy models locally and in real time – without cloud latency, shared infrastructure or the risk of sensitive data leaks. NVIDIA NIM microservices then enable the rapid deployment of optimised models as containers for production inference.