Professional NVIDIA GPUs for AI, ML and visualisation

A powerful server-class GPU for AI and professional visualisation, accelerating AI, scientific computing, graphics and video production.

• NVIDIA Blackwell architecture
• 96 GB GDDR7 with ECC
• Memory bandwidth 1,792 GB/s (Server Edition 1,597)
• PCI Express 5.0 x16
• Max power: 600 W (Server Edition configurable from 300 W)
• Read the real-world benchmarks

Available to order

A universal data-centre GPU of the Ada Lovelace generation: acceleration for LLM inference and training, graphics and video, and multimodal generative AI.

• NVIDIA Ada Lovelace architecture
• 48 GB GDDR6 with ECC
• Memory bandwidth up to 864 GB/s
• Max power: 350 W
• Read the real-world benchmarks

Available to order

For generative AI and HPC: 141 GB of HBM3e at 4.8 TB/s gives large language models and scientific computing more capacity and faster memory.

• NVIDIA Hopper architecture
• 141 GB HBM3e memory
• Memory bandwidth 4.8 TB/s
• Max power: up to 600 W
• Read the real-world benchmarks

Available to order

Blackwell with 84 GB of GDDR7, between the RTX PRO 5000 and the RTX PRO 6000: room for larger models, fine-tuning and visualisation in workstations.

• NVIDIA Blackwell architecture
• 84 GB GDDR7 with ECC
• Memory bandwidth 1,398 GB/s
• PCI Express 5.0 x16
• Max power: up to 600 W
• MIG: one 84 GB or two 42 GB instances

NVIDIA specifications preliminary

Blackwell with 48 or 72 GB of GDDR7: room for larger models, fine-tuning and demanding visualisation, at 300 W in a dual-slot card.

• NVIDIA Blackwell architecture
• 48 or 72 GB GDDR7 with ECC
• Memory bandwidth 1,344 GB/s
• PCI Express 5.0 x16
• Max power: 300 W
• Read the real-world benchmarks

Available to order

The mid-range Blackwell card: 32 GB of GDDR7 for AI inference, design and visualisation, at 200 W in a dual-slot card.

• NVIDIA Blackwell architecture
• 32 GB GDDR7 with ECC
• Memory bandwidth 896 GB/s
• PCI Express 5.0 x16
• Max power: 200 W

Available to order

The Blackwell card for rack servers: 32 GB of GDDR7 for inference and virtual workstations, passively cooled at 165 W in a single-slot, full-height card.

• NVIDIA Blackwell architecture
• 32 GB GDDR7 with ECC
• Memory bandwidth 800 GB/s
• PCI Express 5.0 x16
• Max power: 165 W
• vGPU from release 20.0, MIG up to two 16 GB instances

Available to order

24 GB of GDDR7 in a single-slot card at 145 W: Blackwell for AI, design and visualisation in compact and multi-GPU workstations.

• NVIDIA Blackwell architecture
• 24 GB GDDR7 with ECC
• Memory bandwidth 672 GB/s
• Max power: 145 W
• Single-slot design

Available to order

An energy-efficient data-centre GPU for inference, video and graphics: 24 GB in a low-profile, single-slot card at 72 W.

• NVIDIA Ada Lovelace architecture
• 24 GB GDDR6 with ECC
• Memory bandwidth 300 GB/s
• Max power: 72 W
• Low-profile, single-slot, passive cooling

Available to order

Next-generation RT, Tensor and CUDA cores with 24 GB of GDDR7, in a compact low-profile card at 70 W.

• NVIDIA Blackwell architecture
• 24 GB GDDR7 with ECC
• Max power: 70 W
• Low-profile, dual-slot design

Available to order

Blackwell in a compact, energy-efficient design: 16 GB of GDDR7 for AI workflows and professional graphics at 70 W.

• NVIDIA Blackwell architecture
• 16 GB GDDR7 with ECC
• Max power: 70 W
• Energy-efficient design

Available to order

The workstation generation before Blackwell, in seven cards: RTX 6000 Ada, RTX 5880 Ada, RTX 5000 Ada, RTX 4500 Ada, RTX 4000 Ada, RTX 4000 SFF Ada and RTX 2000 Ada.

• NVIDIA Ada Lovelace architecture
• 16 to 48 GB GDDR6 with ECC
• PCI Express 4.0 x16 (RTX 2000 Ada: x8)
• Max power: 70 to 300 W
• Single-slot, dual-slot and low-profile cards

Available to order

A new class of computer for AI development: prototype, fine-tune and run inference on large models locally, up to 1 petaFLOP of FP4 compute with sparsity in a compact form factor.

• NVIDIA Grace Blackwell architecture
• DGX OS operating system
• 128 GB LPDDR5x unified memory
• Read the real-world benchmarks

Available to order

NVIDIA’s deskside system on the GB300 superchip, built in their own versions by the manufacturers NVIDIA lists. Before we can quote, the distribution asks us for the company, the project and the intended use.

• 1 × Blackwell Ultra GPU, 72-core Grace CPU
• 748 GB coherent memory
• 252 GB HBM3e at 7.1 TB/s for the GPU
• Power: 1,600 W (whole system)
• Compare the alternatives with PCIe cards

Supplied per project

NVIDIA’s 10U system for training and inference: eight Blackwell Ultra GPUs in one NVLink domain. Before we can quote, the distribution asks us for the company, the project and the intended use.

• 8 × Blackwell Ultra SXM GPUs
• 2.1 TB HBM3e in total
• 14.4 TB/s aggregate NVLink bandwidth
• 10U, 15.1 kW (power-supply version)

Supplied per project