The Galapagos Multi-FPGA Architecture

Add to your list(s) Download to your calendar using vCal

• Camilo Vega (University of Toronto)
• Monday 13 May 2024, 13:00-14:00
• Department of Electrical and Electronic Engineering, Room 909B.

If you have a question about this talk, please contact George A Constantinides.

With LLMs and other similar algorithms, there is a tendency to scale to ever larger cluster sizes working on a single application. Thousands of GPU devices, for instance, are run together to perform LLM training. However, FPG As lack the infrastructure necessary for abstraction into the multi-device domain often serving as accelerators to a single CPU device and being constrained to only run applications small enough to fit in a single device.

The Galapagos system we built provides an abstraction layer that allows the user to create large dataflow maps of AXI -Stream connected kernels, so large in fact that it requires the area of multiple FPG As to fit, and it automatically partitions and creates the device interconnection pathways to allow such an application to run seamlessly on an ethernet network connected cluster consisting of many FPG As. Furthermore, our current research is allowing these FPGA clusters to scale better, towards achieving virtual FPG As thousands of times larger than any on the market, and we are creating control infrastructures and other abstractions to retain the simple virtual notion that the user is operating on a single huge device and not on a large network connected cluster. All while retaining a high throughput in the range of 100Gbps and maintaining low latency.

In this presentation, I will describe the philosophy and goals of our open-source platform, Galapagos, showcasing what we have built so far, and what we envision going forward. We will also explore how this platform can expand the capabilities of FPG As, allowing one to embark on exciting new applications as the maximum area constraints are relaxed.

This talk is part of the CAS Talks series.

This talk is included in these lists:

• Andrea Picciau's list
• CAS Talks
• Circuits and Systems Group: Internal Seminars
• Department of Electrical and Electronic Engineering, Room 909B

Note that ex-directory lists are not shown.