What is a transistor?

This is very oversimplified! In the context of DRAM (which is usually the kind of memory that is used), a transistor is a device in memory that lets currents of voltage flow through and charge a capacitor, which is also in memory. A transistor + capacitor is basically how data is stored in memory.

Everything, like pictures, gets converted to a series of 1s and 0s (for example, your phone's wallpaper = 11101). These numbers are represented by electricity. A certain amount of charge (voltage) represents a 1, and this information is stored inside the capacitor.

The problem is that every time you want to read or write to the capacitor, you have to activate the memory cell. When you do this, you lose the charge, so many methods to ensure charge retention have been developed.

What is an FeFET?

An FeFET stands for ferroelectric field-effect transistor, which is a type of transistor that uses ferroelectric material to control current flow. FeFETs are known for their low power consumption and high speed, making them suitable for memory and logic devices.

What's very important for our research is that because of its material, you can program an FeFET to store charge WITHIN the transistor (through polarization in the ferroelectric layer, but this is tedious info)! So now, both the capacitor AND transistor can store data.

However, the material is not completely durable. Many reads or writes to this transistor will wear it down. These are not the transistors currently used in industry, but much research is being conducted to gauge the pros and cons of potential implementation.

This research project did not use FeFET transistors, but instead proposed a hardware-software co-design whose architecture exposed new instructions to the OS for computation directly in memory. Basically... they changed the hardware to make in-memory computation faster and then tweaked the compiler to indicate these changes (yay)!

On a hardware level, they used sense amplifiers for computation along with optimizing how rows were activated (multiple rows at once and only activating necessary columns) to make computations more efficient.

This is important because the data stays closer to the memory where it is stored, whereas transferring data from memory to another part of the hardware (which could be farther and thus more expensive and time-consuming) for computations is less efficient.

We thought that if we added an FeFET transistor to the hardware architecture they used in MIMDRAM, we could further optimize the in-memory computing developments. We looked into applications as to why this would matter, and thus researched HTAP databases.

HTAP stands for hybrid transaction and analytical processing, which is a type of database that allows for both transactional and analytical processing in a single system.

This is essentially for very large databases that are being updated frequently. Not only that, but this data, which is updating all the time, is being used for computation for some sort of analysis.

This is important because it allows for real-time data processing and analysis, which is crucial for applications such as credit card fraud detection. Every time you use your credit card (an example of data that needs frequent updating), the system needs to process that transaction in real-time to detect fraudulent activity (the analytical part).

For example, if you make a transaction in Australia 30 minutes after trying to make a transaction in France (where you reside), an algorithm will take this new information, calculate how likely it is to be fraudulent, and accept or deny the transaction.

Finally, we explored implementing content addressable memory (CAM) with FeFETs for FPGAs within the field of high-frequency trading.

An FPGA is essentially a very simple circuit board (TECHNICALLY... a reconfigurable chip that you can program) that is highly optimized for a specific task (not exclusive to one but for simplicity sake, I am going to say one). So, it does one thing and it does it REALLY well (W - Unix philosophy mentioned)... Essentially the purpose of FPGA is: speed >> anything else

We looked into algorithms commonly used in high-frequency trading, which is a type of trading that involves executing a large number of orders at extremely high speeds. We then developed 4 architecture proposals for the order book (a stage in its algorithm) to improve speed.

With this, every nanosecond counts!

Overall, this summer was beyond transformative. This slideshow only touches on the many fields and subjects we focused on (and that I learned about through the process).

It doesn't even delve into the coworkers I met (fellow PhD students—I got to watch my first PhD defense!), explorations in a foreign country outside of work (since we were able to have funding to do this abroad!), and much more!