XClose

Advanced Research Computing

Home
Menu

Here now - new platforms!

Project Contender Systems

UCL’s Centre for Advanced Research Computing (ARC) has recently installed various new one-off systems mostly with new, or currently esoteric, accelerator cards, which we are promoting as Project Contender. These systems are: 

“mandelbrot” –  
a Graphcore POD16  

OPERATIONAL

 

You may have heard about this one before as it is already up and running. Graphcore is a British contender in the lucrative machine learning market. Their unusual processors are designed specifically for machine learning workloads and can be easily programmed in Pytorch and TensorFlow.  

IPU-POD16 Direct Attach Datasheet — IPU-POD16 Direct Attach Datasheet (graphcore.ai)

“zeno” –  
Qualcomm Cloud AI100 

OPERATIONAL

This has Qualcomm GPUs that are particularly designed for the inference stage in machine learning.  

Qualcomm Cloud AI 100 | AI Inference Processor for Datacenters | Qualcomm

“cricket” –  
ARM CPU, Nvidia GPU 

OPERATIONAL

This has a couple of the popular Nvidia A100 GPUs, but the twist is that the host has an ARM architecture Ampere Altra CPU. 

G242-P34 (rev. 100) | GPU Servers - GIGABYTE Global
NVIDIA A100 | Tensor Core GPU

“locust” –  
Nvidia Grace-Hopper superchip 

OPERATIONAL

We are very pleased to have been offered a pre-release sample of this machine by Nvidia. Generally, it is like the system above in that it has an ARM CPU, this time from Nvidia, and an Nvidia GPU, the new Hopper, but these are linked by very high-speed interconnect and have unified memory space comprising both the CPU’s and the GPU’s RAM so that data transfer between the two is handled automatically by hardware. 

NVIDIA Grace Hopper Superchip | NVIDIA

“terrarium” –  
NEC SX Aurora TSUBASA B300-8 

OPERATIONAL

This is somewhat different. It is described as a vector processor. It has a long legacy with one of its predecessors having topped the Top500 at the Earth Simulator Center. 

B300-8 (nec.com)

"vindicator"

OPERATIONAL

An actual IBM mainframe, so suited to running multiple concurrent applications securely at high transaction rates, high speed I/O, responsive machine learning inference.

IBM LinuxOne

Nvidia BlueField cards 

A couple of these systems also have BlueField cards. These network cards have an integrated ARM architecture Linux computer embedded on the card. The aim is to offload network functions from the CPU to this auxiliary processor or even to work as a supervisor for running virtual machines on the host.  

Nvidia BlueField Networking Platform 

 

We envisage that these will attract users from across UCL. Perhaps the most common use might be to benchmark your machine learning codes on these against your usual current resources. Other uses would include working on those or other codes to see how the novel systems may be leveraged, or power consumption comparisons.  

If you would like access to these systems get in contact soon with ARC’s James Legg (j.legg@ucl.ac.uk) giving a brief statement of what you would like to use them for. Please note that these are experimental machines and so are for projects that concern the machines themselves or their software and are not intended for big production runs.