Alumna Sophie Wilson (SE 1976) has had an illustrious career in computing science. The chip she devised is now in the overwhelming majority of smartphones worldwide, and she is seen as one of the most important figures in her field with a range of lifetime achievement awards and a fellowship of the Royal Society. Writing exclusively for this magazine, she reveals what’s coming next in technology.

Sophie Wilson CV
Born: June 1957
Education: Harrogate Grammar School, Selwyn College 1976–1979 Computer Science
1979
Started at Acorn Computers
1981
Developed the BBC Micro computer
1985
ARM chip entered production
2012
ARM is the most used mobile and computer architecture
1999–2001
Co-Founder & Chief Architect at Element 14, which was acquired by Broadcom in 2001
2001
Research Fellow and Director at Broadcom Corporation
2019
Appointed a Commander of the Order of the British Empire (CBE) for services to computing
I last wrote for this magazine in autumn 2014, predicting that the future of computing lay in power-efficient microprocessors built on a single silicon die, all specialised for particular tasks. Over the past decade, all that has happened and more. Thanks to modern microprocessors, we now have accessible ‘AI chatbots’ on almost every website, with ‘AI image generation’ at your fingertips. What is next?
Firstly, hardware will continue to evolve. Machine learning is coming to your computers. Machine learning is one of the basic components of AI; in essence, it is a system that can learn and carry out tasks without human input. Indeed, it is already here — speech recognition is now done with machine learning algorithms, and so is unlocking your phone or computer with your fingerprint or face. You may not think of those things as being similar to ChatGPT, but they use similar algorithms to what ChatGPT uses and they are processed locally on processors on your device. In fact, if you bought a device or computer in the last four or five years, then you can run “proper” machine learning on it, which is usually more power efficient than sending all your data to a supercomputer in the cloud.

So where do supercomputers fit in with the future of AI? While modern computers are built out of a few silicon chips which contain one or more silicon dies with multiple processors on them, supercomputers are built out of lots of silicon chips. Furthermore, supercomputers purpose-built for machine learning tend to include many more highperformance processors on the chips, which are also highly specialised. For example, supercomputers have specialised processors called Neural Processing Units (NPU). NPUs are a relatively new type of processor that can compute similarly to our brains, which makes them useful for developing artificial intelligence. Over the last decade, there has been a dramatic race in NPU performance — a gain of 1000x for machine learning algorithms on a single chip — the same 1000x gain for Central Processing Units (CPU: the main processor in a computer) took three times longer. In fact, they have already trickled down to your devices: the machine learning algorithms for speech recognition and unlocking your phone are handled by versions of NPUs on your phone. So, there’s technically nothing that a supercomputer does that an ordinary computer, or your tablet or phone, can’t do — a supercomputer just does it a lot faster.
Measuring this sort of computer performance has always been tricky. With machine learning it’s even more tricky, as performance becomes more about how a processor on the chip called the General Purpose Graphics Processing Unit (GPGPU) can handle specific algorithms. We currently measure their performance scale in Tera-Operations Per Second (TOPS), with an operation defined as a simple task like a multiply or add.
For context, a basic modern computer chip like an Apple M3 chip (available on Apple’s latest laptops) is capable of 18 TOPS running on just a few watts of power. An Intel Core Ultra ‘Meteor Lake’ chip offers 10 TOPS, an AMD Ryzen ‘Hawk Point’ offers 16 TOPS and their future offerings are likely to be around 40 to 45 TOPS, which is what Microsoft advise for machines to be capable of running their ‘Copilot AI’ services locally. Look out for heavy marketing of ‘AI PC’ this year — far in advance of anyone knowing exactly what these services will provide, what use it is to people and therefore, what they’re worth.
So that’s the hardware evolving — more processors doing more things, just like before. In a way it is properly boring — but boring is good, and it lets engineers work towards the future.

The real fun is in the future of software. The software for machine learning is evolving at pace, moving beyond the sole control of the technical mega-corps (in this case: Microsoft, Google, Amazon, Alibaba, Baidu).
When OpenAI’s machine learning model ChatGPT, a Large Language Model (LLM) to be specific, first appeared it galvanised theMarket and the research establishment. Some companies, such as Meta (formerly Facebook), made their LLMs open source, and other companies also released models for things such as image generation.
This has spawned a cottage industry that is collectively innovating rapidly, especially as these models are generally designed to run on accessible modern computer chips like the Apple M3. Most importantly, there are now models of various sizes.
The size of a Large Language Model is measured in parameters (a technical term for the factors that an AI system learns from), and they get better as their size increases. For example, a 2 billion parameter one is easy to run ‘locally’ on your phone but has noticeably poor performance. By performance, I mean that it is capable of responding to questions/conversations but is a bit dull and limited in its answers and prone to ‘hallucinations’—outright making things up.
At 7 billion parameters, the LLM gets better at doing the tasks you want but is still quite error-prone. You begin to see something useful at 10 billion and 14 billion, but these need a moderately powerful PC to run. For 70 and 140 billion parameter models, you need a top-of-the-line machine with high-powered components. These are still accessible to a consumer, albeit at a steep cost.
I’ve concentrated on LLMs here, but this applies to other uses of machine learning—image generation and image processing of all kinds (increase resolution, remove objects, corrections). And I have to add a warning: LLMs of even the largest size are still prone to hallucinations, and have all kinds of bias. But they’re very useful. For example, I’ve used open-source local LLMs for summarisation, both because I can’t send things to the cloud for security reasons or because it is easier to experiment with local models. On the other hand, I did try fairly hard to make a Christmas card with a robin sitting on a garden fork handle in a snowy garden on a locally running image generation model, but I couldn’t get the result I wanted... I think there’s now a new skill for humans of the future: working out a prompt for a machine learning programme.
One last thought: there are a lot of people are discussing ‘AI’ as a magical cure-all or ultimate threat. The dictionary definition of ‘intelligence’ is “the ability to apply knowledge to manipulate one’s environment” as one of its clauses. The machine learning we have today certainly doesn’t do that. At best AI is only 50% of that — it is definitely artificial. Any manipulation of the environment is left to the user.

Glossary
OpenAI
OpenAI is a relatively new organization that has gained prominence for developing artificial intelligence technologies, most notably ChatGPT.
Large Language Model (LLM)
A Large Language Model (LLM) is a computer program that has been fed enough examples to be able to recognize and interpret human language or other types of complex data. Many LLMs are trained on data that has been gathered from the Internet—thousands or millions of gigabytes’ worth of text.

Selwyn is proud to offer the Sophie Wilson scholarship in Scientific Computing to help educate the next generation of Computer Scientists. This fund, generously funded by Sophie Wilson, will help bring a wider range of students to the field, building a stronger and more inclusive scientific computing community. If you would like to learn more about what impact your donations could have, please contact the Development Office on development@sel.cam.ac.uk.