Wafer-Scale Chip to Power Edinburgh Supercomputer

one month ago by Luke James

Edinburgh’s EPCC (formerly the Edinburgh Parallel Computing Centre) supercomputing facility has recently announced that it will be using the Cerebras CS-1, the world’s fastest artificial intelligence computer processing system, in the Edinburgh International Data Facility.

EPCC is the superconducting centre of the University of Edinburgh, and it is planning to deploy a Cerebras CS-1 supercomputer for its AI-based research.

 

What is Cerebras CS-1?

The Cerebras CS-1 system is an example of a ‘large wafer scale engine’ processor alongside an HPE Superdome Flex Server system for pre-processing and front-end storage. Combined, these two technologies have been claimed to hugely reduce the amount of training time required for AI models.

The CS-1 was announced back in 2019 following three years’ worth of development by Cerebras. Described as a ‘unicorn startup’ by Forbes, the company was co-founded by hardware architect Sean Lie and CEO Andrew Feldman, the former founder and CEO of Sea Micro, a micro-server innovator that was acquired by AMD in 2011.

Cerebras’ approach is simply to build the largest chips ever seen and break down current barriers to speedy computing. In 2020, the company was valued at $1.7 billion and has raised around $200 million from top investors. The CS-1’s European deployment marks a huge stepping stone on the company’s path to market dominance.

 

Cerebras Comes to Europe

The deployment of the Cerebras CS-1 at EPCC will mark the first occurrence of the system being used in Europe. In this case, it will be tasked with supporting research into natural language processing and data science across public and private organisations and academia.

 

The Cerebras CS-1 supercomputer

Image credit: Cerebras

 

The CS-1 WSE (wafer-scale engine) is a massive single wafer processor. It measures 46.2 cm2, which makes it 56 times larger than leading graphics processing units. Cerebras also claims that it has 54 times more cores, 450 times more on-chip memory, and 20,833 times more fabric bandwidth. The chip offers major improvements in data movement efficiency, massive parallelisation, and memory co-located with processing.

Given that building a wafer-scale chip has for quite some time been a goal widely thought of as unreachable by the semiconductor industry, the sheer amount of computing power that is packed onto a single silicon substrate in the CS-1 marks a breakthrough in the industry. 

 

Cerebras CEO’s Vision

According to Andrew Feldman, the CEO and co-founder of Cerebras, the vision with CS-1 was to “reduce the cost of curiosity”.

“We are excited to bring our industry-leading CS-1 AI supercomputer, coupled with HPE’s advanced memory server, to EPCC and the European market to help solve some of today’s most urgent problems,” he said in an announcement.

By building huge chips like the CS-1, Cerebras believes it will be able to store an entire neural network on a single device, eliminating the need to scale across multiple devices and memory layers.

Said Feldman: “We look forward to the myriad experiments and world-changing solutions that will emerge from EPCC’s regional data centre”.

Comments

  • Tom Wlazlak - 2021-05-24

    There was a question asked years ago, Why can’t permanent magnetic force be used as an energy source?
    The answer is Permanent magnetic like Gravity does not conform to the laws that energy must follow.
    But water can be used as energy by the use of Gravity to produce electrical power, therefor-
    Permanent magnetic force can use iron in the same way as water to produce electrical power output if used in a type of recycle system motor. Called EttCM Energy technology.
    This technology is a proven method of using ferrous metals as energy to produce usable power output.
    Tom Wlazlak - .(JavaScript must be enabled to view this email address)

    Reply