Creating Optically Interconnected GPUs


Photonics provides a very promising solution to the problems that many high demand microelectronic devices such as CPUs are encountering today. The lithography processes for making CPUs today has enabled the creation of devices where billions of transistors can be shrunk into a small 2×2 enclosure. This however, allows wires to be placed so close together that quantum tunneling and the production of large amounts of heat become very concerning problems for giant processor manufacturers like Intel. Photonics provides a solution to these problems by allowing information to be transmitted across a chip using optical waveguides, which requires less energy, and thus generate less heat.


However, almost all electronic devices today use silicon based technology; from cell phones, TV sets, or laptops, to over half of the medical equipment used in hospitals. Silicon has be used since the mid-1950s to manufacture computer components, and much work has been done to bring the rapid manufacturing capability  and reliability of silicon Integrated Circuits (ICs) to the point that it is at today. It would not be a realistic objective to switch an existing silicon architecture to a purely photonics based technology.

An alternative to switching to a complete optical chip architecture would be to create a hybrid integrated circuit by integrating photonic circuits with existing silicon technology. Some researchers – namely Milos Popovic from MIT– have proved that this is possible by using micro-ring resonators as the photonic alternative to a transistor in his design and by utilizing optical waveguides as the medium for inter component communication within a RISC-V architecture based microprocessor. The implementation was shown to be successful and serves as compelling evidence that photonic circuits are indeed power efficient and fast. More detailed information about Milos Popovic’s implementation can be found in his publication “Single-chip microprocessor that communicates directly using light.”


However, an inevitable problem that manufacturers have acknowledged lies within the low popularity that this technology currently has. Photonic circuits have proven to be more energy efficient than their state of the art silicon counterparts, but they are simply not used in any mainstream devices. In addition, since it is a very new technology, it cannot currently compete with silicon technology on its own.  With that said, it will be difficult to generate a large enough interest and consumer base to make this a worthwhile technology.

An important but overlooked fact, however, is that many modern computing devices such as desktop computers and mobile phones have a display powered by a Graphical Processing Unit (GPU). In fact, the first thing an individual will notice when then he or she power on their device is this display. Current silicon based GPUs can be very power hungry and often generate an excessive amount of energy.

To address this issue and to generate public interest to photonic technology we propose to implement a hybridized photonic-silicon version of a popular, pre-existing Graphic Processing Unit to demonstrate that high power efficiency as well as an increase in performance (reduced data latency) can be achieved by utilizing photonic integrated circuits to allow for the intercommunication between the silicon based components within a single chip(IC).

Some Useful Definitions:

Optical Wave Guide – A transparent tube usually made with a polymer with a high refractive index. This allows light entering the tube to travel and bend along its length and exit out the other end. GPU – Graphical Processing Unit. This is the component responsible for generating the pixel data for computer screens.

CMOS – Complementary Metal Oxide Semiconductor. A technology used to construct basic elements inside of an integrated circuit, such as transistors or diodes. Attenuation – The amount by which a signal loses its original power when traveling through some medium. Photonics – An area of science that deals with the generation, manipulation, and detection of light.

Silicon Germanium – A material used in semiconductors to allow for low current and high frequency performance.

Silicon Micro-Ring Resonator – The basic unit of the photonic circuit. When the voltage across a  micro-ring resonator is positive, the resonant frequency of the micro-ring changes such that it resonates with the incoming light source and allows photons to pass through it.

Electro-Optic Modulation: The method in which the resonant wavelength of the micro-ring resonators is changed. In our case, this is achieved by applying a voltage across the micro-ring resonator.

Quality Factor: A measure of how underdamped an oscillator or resonator is – higher Q, lower rate of energy loss

Decibel – A logarithmic unit to express a ratio of two physical quantities. In this case,

SOI – silicon on insulator – A technology used in microelectronics to reduce parasitic device capacitance

Design Approach and Type of Design Used

Our goal will be to fabricate a simple open-source Graphical Processing Unit (GPU) using the methods similar to those used in creating an optical CPU. Graphical Processing Units are everywhere in modern day productivity machines. The video output quality attainable today would nto be possible without the use of GPUs. For instance, generating data for thousands of pixels on a computer screen for a graphic intensive video game is simply a too demanding task for many mid-grade computers. In addition, it is a very power hungry task due to the fundamental architecture of most GPUs.

Photonic integration to silicon electronics is a relatively new concept, and thus requires smaller to prove that it could be considered useful and competitive in today’s market. Our research group will aim to create a more energy efficient, photonic interconnected GPU that can be used to carry out simple graphical processing tasks such as the rendering of a 3D model. To validate and test our hypothesis that optics is indeed a practical solution to producing more energy efficient, higher performance devices, we will compare the regular silicon implementation of a our leveraged graphical processing unit to the performance of our photonically interconnected GPU, performing the same operations on both devices, under very similar external conditions (temperature, location, etc.).

Role of the Researcher

The researcher’s role will be to study photonic circuit design, and to develop an photonic-silicon device to allow optical communication between different components within an integrated circuit (IC) interconnection. This will involve the design of non-linear circuit components, and the selection of appropriate parameters of those components to adequately meet the specifications for our design.

Essentially, the researcher would look at Verilog code from the MIAOW open source GPU project , run that code on a Xilinx FPGA where the code will be turned into a hardware implementation (because a Xilinx MicroBlaze Soft CPU was used in the creation of the opensource GPU) see where the photonic-silicon interface can be made, and design those accordingly. Once that is complete, researchers will finalize, designs, move onto fabrication of the IC, and then test the validity of the design. (This a link to the repository of code:



This is a process that we predict will take approximately 14 months. The general procedure will be as follows:

(1) Obtain the MIAOW open GPU project from the owners at University of Wisconsin-Madison

(2) Analyze the design of the circuit

(3) Put design on FPGA

(4) Manufacture open GPU design in silicon

(5) Perform benchmarks on this GPU and implement a motherboard-like controller on the FPGA

(6) Form the list of operations we will perform to effectively measure and show off the performance of the GPU

(7) Study optical circuit design

(8) Integrate photonics into this GPU

Using methods learned from Professor Milos Popovic and his implementation of an optically interconnected RISC architecture CPU.

(9) Manufacture it with optical waveguides

(10) Perform benchmarks




Data Collection and Analysis

We will capture images of our photonic-silicon ICs to clearly convey our final implementations and designs to students, prospective researchers, and other interested parties. We will first collect data of the rendering performance of the regular silicon implementation of the GPU using open source benchmarking software, and then compare those results to our photonic-silicon implementation of the GPU.


There are no deep ethical concerns surrounding our research. But we will make it very clear to our lab participants, that all figures must be accurately recorded, and that no figures have been exaggerated. In addition, we will ensure that each of the resources that we will be using will have been bought using money granted to use through the foundations we seek funding from (Universities such as CU Boulder, MIT, UC Berkeley, and other participating universities), the software or source code we will use will use will be licensed to us, or will be obtained with the owner’s permission. In addition, we will also give proper and equal acknowledgement to the contributors to everyone who participates in our research project as well as those who have provided us with any licensed software.

Reliability and Validity of Methods and Results

We have already seen that the methods used in fabricating the optical CPU were successful. In theory, transferring those methods over to fabricating a GPU, should also work and yield results of similar characteristics in reporting. We will be using standardized methods for measuring the Input/Output (I/O) performance and throughput of this GPU. These will all be measurements which are popular in usage. Many benchmarking tools have been created to carry out this procedure, and we will leverage those programs to provide us with data. However, in some instances we will have to make our own software, to perform this task. Usually the data recorded will be the times in which it takes for a certain to task to begin and end. This only involves counting the number of system clock ticks; therefore the creation of a reliable benchmark software can be done. Of course, there will be a calculated margin of error, and that can be found by asking the manufacturers of the equipment we will use.


Quarter 1

(1) Obtain the MIAOW open GPU project from the owners at University of Wisconsin-Madison


(2) Analyze the design


(3) Put design on FPGA


(4) Manufacture open GPU design in silicon


Quarter 2

(5) Perform benchmarks on this GPU and implement a motherboard-like controller on the FPGA


(6) Form the list of operations we will perform to effectively measure and show off the performance of the GPU


Quarter 3

(7) Study optical circuit design


(8) Integrate photonic circuits with silicon and Validate Designs

March June

Quarter 4

(9) Manufacture GPU with optical waveguides

July – August

(10) Perform benchmarks

September – October


Resources and Materials, Budgets, Costs

We plan to start with only 4 lab assistants and accommodate each of them with their own laptop computers for research purposes. As we move along our development path, we will need to purchase our own Field Programmable Gate Arrays to act as the motherboard for our GPU to interface with, we will need to rent our own chip fabrication equipment, purchase licenses for our Photonic-Silicon integrated circuit (IC) design software, and will need to also pay our lab assistants for aiding our research.

Computers x 4 4 x $499 = $1,996
Fab Equipment Rental x 1 $10,999
FPGAs x 2 2x $129 = $258
Fabricating PCBs $200 x 4 =$ 800
Design Software $1,000
Benchmark Software $159
Lab Assistants x 4 4 x $2,000 = $8,000
Total $23,212



We are limited in how powerful our GPU will be because the open source GPU we have selected is based on a limited version of an AMD GPU architecture from 2011. In addition, because of the large expense of fabricating the photonic-silicon ICs, not very many mistakes can be made. Otherwise, we risk the event of wasting a large amount of money on already defective ICs.

We are also limited by the fact that we will not be creating a piece of hardware that will be recognized by many computers. There will be a custom interface for connecting the GPU to it’s host device (the FPGA). We will also be creating the drivers necessary (ie define the communication protocol) to communicate with the GPU, load it with programs, and give it commands.



GPUs are in all devices that output some sort of display. Especially in desktop computers, laptops, cell phones, or even smart watches. This experiment will be limited in that we will only be measuring the render time and performance of a predetermined, simple 3D model. This GPU will not be used to play video games of any kind (although maybe it may have the potential to play very low graphic games), it will not be used to show movies. This experiment is to serve as a further proof of the concept of the possibility of integrating silicon devices, with photonic devices.

In terms of the hardware within the IC, our micro-ring resonators (the on –off switch of the resonators) will be very sensitive to changes in temperature and even the small expansions and contractions created in the layering of the silicon in the chip itself. This causes swings in the resonant frequency of the micro-ring. If the micro ring does not have the desired resonant frequency, this will cause undesired effects such as the transmission of photons when not desired, or vice versa. This forces our research group to construct control circuitry, and this will add upon the time required to carry out this experimental GPU design. However, the

Final Product

The final outcome of this product should be a lower power graphic processor which uses only silicon germanium optical waveguides to communicate between the main onboard memory and the processing core(s) of the GPU itself. We expect to see a decrease in power consumption (due to the fact that only a single laser at a given wavelength of light will be used to generate the light that will pass through all transistors on the IC) as well as an increase in GPU IO speeds.


Jump to annotated bibliography.

Jump to lexicon.