Week 3: Graphics Cards and Spectrometer

This week we spent our time in the lab using the spectrometer. We familiarized ourselves with starting up the Menlo Systems laser and sending the signal from the 50/50 coupler to the spectrometer. We practiced using the various functions of the instrument such as re-centering the wavelengths and bounds, adjusting the intervals of sampling and changing the resolution.

In addition to using the spectrometer, we tested both Python and MATLAB code on the secondary desktop, which housed a low-end graphics card. This was done to see if the GPU would take the load off the CPU. Using the activity monitor tool, we noticed a difference in CPU and GPU percent usage. The Python code ran purely with the CPU while the MATLAB code utilized the GPU. Because of this, the MATLAB code ran faster and smoother. Since the Python code did not use the graphics card, we would need to code the GPU to handle the load. More tests with the code will be run in the coming week.

Next, we began to research graphics cards to determine the best price-to-performance ratio. Table 1 shows the specifications of top of the line graphics cards that are within the allotted budget of $1000. We looked towards other scholarly imaging projects that utilized graphics cards to get an idea of the criteria that the GPU needs to meet in order to be effective for the project. Additionally, we began to familiarize ourselves with the different architecture of the GPUs manufactured by NVIDIA and AMD.

Table 1: GPU Comparisons

From our understanding, core count and clock speed are more important criteria, as each card has plenty of memory. Based on OTC projects that utilize GPUs in their system, most cards used only had approximately 3GB - 4GB of memory, while the cards we are considering all have 8GB. It is recommended that a desktop’s power supply be able to supply approximately 550W - 650W, depending on the card. In addition to the PSU, the dimensions of each card could rule them out of the list if there is not enough space in the desktop to mount the GPU.

In addition to the numbers specifications, each brand and generation of graphics cards have different architectures. NVIDIA’s new Turing architecture is more of a multi-purpose GPU architecture than the previous architecture, Pascal. It can perform Pixel Processing, Artificial Intelligence, and Real-Time Ray Tracing all at the same time. With this in mind, it may be possible that this new technology may be more beneficial to this project than having a pure core count and processing speed increase. This means that any RTX card could make a great impact regardless of the version, however, more research needs to be done on this topic. In contrast, AMD’s RDNA architecture is hyper-efficient with less latency, power draw, and bandwidth than its predecessor. AMD GPUs are built for power efficiency and affordability while still delivering high performance. With this in mind, AMD graphics cards could be the preferred option if the PSU of the desktop does not output enough watts to power the generally more power-hungry RTX cards. With these considerations in mind, we believe the RTX cards are currently the better option as they generally have more pure processing power and its other functions may prove useful in the future.

Our plans for next week are to dive deeper into NVIDIA’s Turing GPU architecture through their official technical papers and to be able to run the system confidently to produce a visual output. We plan to narrow down our GPU options by examining the constraints of the desktop, such as size limitations and GPU form factors as well as if there is enough power to supply the added GPU. In addition to that, we plan to determine if there could possibly be any other bottlenecks of each desktop. One desktop has a more powerful i5-6500 CPU, but a tighter size constraint due to the chassis while the other has a weaker CPU from Intel’s 2nd-generation and a larger chassis.