HIGH PERFORMANCE COMPUTING
“AI and HPC end-use cases are seeing an increasing trend towards demographic data, because now we can collect more data, data is becoming more complex, and compute nodes are becoming more powerful”
and complexity of data are increasing and models are getting larger it means that one, two or three GPUs in a system will no longer be enough for AI training. You need six, seven, eight or even 20 in a single system. If you go on like this, there’s no limit, there’s no end to adding more GPU cards. So how do we solve the problem? One aspect of this topic is physicists
creating new semiconductor materials that are more capable of retaining data and transmitting data faster in large quantities. These problems will be solved in the next five years. But as you know, the IT industry moves fast, and no one has time to wait five years.
So an immediate solution to this is
launch the workload on the entire cluster. This is my opinion, this doesn’t
represent Gigabyte, but I think that you will see more and more granular hardware designs into GPU technologies that will allow for virtualisation and also physical isolation of GPU compute units within the GPU device. For example, if I take a GPU card, when we used to virtualise the GPU card it was the entire device. Because the GPU architecture is the same for all compute units inside, you can do mostly the same kind of workloads. It doesn’t matter which virtual machine
you use, it doesn’t matter which piece of hardware you get, because the GPU hardware is the same. The compute units inside are the same. What we’re going to see is a GPU architecture trend with different compute units of different sizes and different types of memory management within the same GPU card. This means that you can isolate the GPU hardware, so that you can allocate
www.scientific-computing.com | @scwmagazine
different parts of the same GPU to different kinds of users, running different kinds of data analysis.
What are the challenges facing HPC and AI hardware development? There are really two aspects to this question. The first aspect is the physical limitation. As you know, now, most of our semiconductors are based on silicon. So in terms of physics, the existing semiconductor materials are not capable of supporting any more downsizing in the manufacturing process. There’s a limit to that. You can play with semiconductor architecture in terms of the transistor channel layout, voltage control, conducting and nonconducting channel width – you can play with all those things, but in the end, there’s a physical limitation. The reason I talk about this is that as you
go into the AI, data will get larger and when it gets larger, the AI models, the neural networks, will become larger. So if the size
the second aspect of the topic, which is leading to immediate changes in the CPU and GPU architectures in the next two years, in 2022 and 2023. What do I mean by CPU and GPU architecture changes? Data is getting heavier, larger, and therefore you need more memory, faster compute, minimised latency in data transfer, and also compute speed. And with the challenges that we’re facing now, how do you do that? Intel and AMD are proposing a hybrid CPU architecture design, meaning the cores will change to a hybrid design in the sense that not all the cores will be the same size. What I can tell you is the next generation of CPUs will have much faster, much larger cache memory. Just that design change will solve many of our current challenges.
Hsueh-Li Wang is the EMEA business development manager for HPC, AI, virtualisation server solutions, data centre projects at Gigabyte, which has 30 years of engineering expertise in the development and production of server solutions covering a myriad of uses.
Autumn 2021 Scientific Computing World
9
GarryKillian/
Shutterstock.com
Page 1 |
Page 2 |
Page 3 |
Page 4 |
Page 5 |
Page 6 |
Page 7 |
Page 8 |
Page 9 |
Page 10 |
Page 11 |
Page 12 |
Page 13 |
Page 14 |
Page 15 |
Page 16 |
Page 17 |
Page 18 |
Page 19 |
Page 20 |
Page 21 |
Page 22 |
Page 23 |
Page 24 |
Page 25 |
Page 26 |
Page 27 |
Page 28 |
Page 29 |
Page 30 |
Page 31 |
Page 32 |
Page 33 |
Page 34
