Nvidia - Petaflops

NVIDIA A100 TENSOR CORE GPU

These are the flagship Nvidia GPU for HPC and AI applications including Deep Learning and the largest simulations.

These are the first cards with Nvidia’s new Tensor Cores featuring Double Precision Tensor Float (TF32) for 100% performance improvement in some double-precision simulations

The A100 also supports MIG or multi-instance GPU allowing each GPU to function as 7 separate GPUs which has been shown to increase inference speeds by up to 700%.

Card Series	Ampere
CUDA Compute Cores	6912
AI Specific Tensor Cores	432
Ray Tracing RTX Cores	0
Bus	PCIE4 and HGX (SXM3)
RAM	40GB HBM2
Ports	N/A

We recommend these cards for the highest performance in a mature data centre HPC and AI environment. These are available as PCIE4 cards or for highest performance in specialised HGX bus computers supporting the new 4 GPU and 8 GPU / 1Petaflop configurations and even Nvidia 16 GPU / 2 Petaflop systems providing up to 9.6TB/sec bandwidth.

NVIDIA V100 - The first Tensor Core GPU

These Tesla series cards are fast PCIE3 and SXM2 optimised for High Performance HPC, AI, Simulation, and Render farm applications.

Card Series	Tesla
CUDA Compute Cores	5120
AI Specific Tensor Cores	640
Ray Tracing RTX Cores	0
Bus	PCIE3 and SXM2
RAM	32GB HDM2 (a 16GB model also available)
Ports	N/A

We recommend these cards for most HPC, AI and rendering. These cards are available in PCIE3 or in specialised 4 and 8 GPU systems from a number of vendors. The V100 is a great and proven choice for most AI, HPC, Simulation and Render environments wanting high performance.

NVIDIA QUADRO RTX 8000 - The World's First Ray Tracing GPU

These are a great multi-purpose GPU with extremely high performance.

These cards combine CUDA compute cores, AI specific Tensor cores and Ray tracing cores. They also have 4 display ports and a VR specific virtual link port. They perform well with anything from AI to rendering to virtual desktops.

Card Series	Turing
CUDA Compute Cores	4608
AI Specific Tensor Cores	576
Ray Tracing RTX Cores	72
Bus	PCIE3
RAM	48GB GDDR6
Ports	4 Displayport 1 Virtual Link

A maximum of two RTX8000 cards may be linked in one system with a bridge to provide around 100GB/sec of bandwidth.

NVIDIA QUADRO RTX 6000 - The World's First Ray Tracing GPU

These are a great multi-purpose GPU with moderate/high performance. This card is similar to the RTX8000 but limited to 24GB RAM.

Card Series	Turing
CUDA Compute Cores	4608
AI Specific Tensor Cores	576
Ray Tracing RTX Cores	72
Bus	PCIE3
RAM	24GB GDDR6
Ports	4 Displayport 1 Virtual Link

A maximum of two RTX6000 cards may be linked in one system with a bridge to provide around 100GB/sec of bandwidth.

NVIDIA QUADRO RTX 5000 - The World's First Ray Tracing GPU

These are a great multi-purpose GPU with moderate performance. This card is similar to the RTX8000 and RTX6000 but limited to 16GB RAM and with a lower number of cores and bandwidth and targeted at design and visualisation.

These cards combine CUDA compute cores, AI specific Tensor cores and Ray tracing cores. They also have 4 displayports and a VR specific virtual link port. They perform well with rendering and visualisation workloads as well as virtual reality.

Card Series	Turing
CUDA Compute Cores	3072
AI Specific Tensor Cores	384
Ray Tracing RTX Cores	48
Bus	PCIE3
RAM	16GB GDDR6
Ports	4 Displayport 1 Virtual Link

A maximum of two RTX5000 cards may be linked in one system with a bridge to provide around 50GB/sec of bandwidth.

NVIDIA T4 - Flexible Design, breakthrough Performance

These are a great multi-purpose GPU designed for industry standard servers. The card draws only 70W power and is single slot.

These cards combine CUDA compute cores, AI specific Tensor cores and Ray tracing cores. They are recommended for Datacentre use including Deep Learning, HPC, Video transcoding, Virtual Desktops and 3D Rendering.

Card Series	Turing
CUDA Compute Cores	2560
AI Specific Tensor Cores	320
Ray Tracing RTX Cores	40
Bus	PCIE3
RAM	16GB GDDR6
Ports	NA

Multiple cards may be installed but are limited to 32GB/sec transfer.

We sell the card separately and in systems from the leading hardware vendors.