We are very happy to already welcome you to May’s edition of PyData Trójmiasto hosted by Graphcore!
[Registration info – important!]
Number of seats is limited to 50. Please provide your full first and last name while registering for the event, here on meetup. The list of attendees freezes 4 hours before the event starts – in case of any urgent changes please leave us a note via firstname.lastname@example.org . Don’t forget to bring your ID for the security check.
18:00 – 18:05 – Meeting boarding
18:05 – 18:10 – A few words about PyData
18:10 – 18:50 – Demystifying AI acceleration by Mateusz Kasprzak
18:50 – 19:30 – Stable Diffusion in practice by Paweł Rościszewski
19:30 – networking & pizza
Demystifying AI acceleration. How does dedicated AI hardware accelerate operations such as convolution, matrix multiplication, normalisation and etc. by Mateusz Kasprzak
A few words from Mateusz
Machine Learning accelerators enabled creation of enormously large, compute intensive AI models that have ground braking capabilities in content creation. Expectations for AI commercial applications are very high, but there is a growing resources problem that has to be addressed. Despite significant progress in hardware performance, every year we observe increase in number of compute units and time required for training . Large Language Models such as GPT-3 are reported to be train on thousands of GPUs for time that exceeds 1 month. To fully understand what is possible to achieve in the near future, it is critical to understand capabilities of hardware that is powering AI revolution. Maintaining current scientific research momentum will require new, much more efficient accelerators, otherwise AI will get monopolized and stale. Join me in „Demystifying AI acceleration” session where I will explain how does AI acceleration hardware work and what are it’s future challenges. I have 7 years of experience working on software for GPUs and ML accelerators. My specialization are computational kernels and I gained experience working in projects like: Intel GPU Drivers, compute Kernels for Intel/Nervana Crest, compute kernels for the Intel OneDNN library, Graphcore Poplar SDK software. I will discuss:
Stable Diffusion in practice: generating music based on natural language descriptions using IPUs in the cloud by Paweł Rosciszewski
Are you also impressed by the capabilities of recent generative AI models? For example, Stable Diffusion generates detailed, high-quality images based on descriptions in natural language. But did you know that generative AI is not limited to text and images? Have a glimpse on this topic through the eyes of a Graphcore AI Engineer and see how these fascinating models can be used for a music generation service supported by Intelligence Processing Units in the cloud.
A few words from Paweł
Since high school I’ve been asking myself what AI really means and how it relates to human consciousness. That is when I borrowed the book „The secret of the Chinese room” by Stanisław Lem. That is also when I implemented the first, naive version of the Gwent card game inside a Witcher-inspired world. Knowing that the dwarves were only shouting random phrases from Andrzej Sapkowski’s books helped me to cope with the anxiety about how dangerous AI can become. Working at Graphcore and being close to the internals of AI models helps me to keep calm. AI gets so good at many things, but how good can it get at creating such a personal and intangible thing as music? After years of lecturing at Gdańsk University of Technology I know that in order to really understand something, you need to try and explain it to others. Come to my first talk about music generation to help me understand what I’m actually doing.