OTC CatchUp #120 Summary

About OTC CatchUp

OTC CatchUps are weekly informal sessions involving project showcases and technical discussions. They are held every Saturday from 10:30 PM IST. Join in!.
For all summaries, please visit catchup.ourtech.community/summary.

OTC CatchUp #120

Date: 25-02-2023

Duration: 4 hrs 12 mins

Topics Discussed

Upgrading SSDs does not necessarily mean better performance since it is dependent on various factors such as processor and motherboard support.
- Jia Harisinghani recently upgraded her SSD and the result was not much noticeable.
  - Pranav Dani suggested she use a Linux distro since it uses fewer resources generally.
Rishit Dagli talked about his recent meeting with Steve Wozniak.
We discussed the application of 3D transforms in computer graphics after Rishit Dagli showcased his project.
- We delved a bit into the math behind the representation of tensors and what texture means. Most of it revolves around Matrices.
- The operations revolving around matrix calculations in his project are performed using GPU.
- We discussed some ways to imagine higher dimensions.
  - Relevant links:
    
    Ways to visualize higher dimensions
    
    Thinking outside the 10-dimensional box - a video by 3Blue1Brown.
Tushar Nankanitalked about the previous instances where we discussed stable diffusion and how it works.
- Stable Diffusion Web is one such example that uses a latent text-to-image diffusion model capable of generating photo-realistic images given any text input.
  - Rishit Dagli suggested that this might be a spinoff and that the real researchers who built it were Stability AI.
Rishit Dagli shared Call For Tiny Papers for ICLR 2023.
- The deadline is 15th March 2023.
- The Call For Tiny Papers is a good way to get started with research.
We discussed two methods in JS - Slice and Splice - for creating a temporary copy of an array.
- This discussion was engendered after Tushar Nankanishowcased jsonxp since we were discussing some ways in which the arrays could be handled.
Google has shut down multiple prestigious coding competitions.
- Affected coding competitions include code jam, kick start, hash code, and more.
- Codejam Program Manager and team was laid off.
- Alphabet fires 100 robots that cleaned cafés, shuts down Everyday Robots project
- Other coding competitions like ICPC and CodeChef are still going strong.
We discussed some ways in which CatchUp Summaries could be automated.
- Google Meet supports native attendance recording but only for the Enterprise or Business users.
We talked about companies like Amazon coming to college with roles providing lower packages.
- While these companies are facing overhiring issues, colleges use the company names in their favor.
Ashwin Kumar Uppalashared that GitHub Campus Expert had some internal security issues recently due to which one of the contractors had complete access to Campus Expert’s portals.
- This led to the removal of most members from the portal, but this issue was soon fixed.
- He also said that he was not able to access raw.githubusercontent.com on Jio network.
  - We have also discussed about it in one of the previous cathups.
Ashwin Kumar Uppalashared that his browser (Safari) was consuming a lot of resources for the meet.
- Browsers might be using more resources due to more available resources.
  - Pranav Dani talked about Parkinson’s law for people to relate with the computer resource issue where the device might use more memory only because it has the luxury to access it.
Ashwin Kumar Uppalatalked about how it is difficult to find good first issues on GitHub since Twitter bots swoop them and it becomes difficult to find and contribute.
- He also shared a blog about - Good First Issues Don’t Exist.
While on the topic of companies like Google shutting down multiple coding competitions, Anil Harwani tried to explain how companies are trying to effectively increase income by reducing expenses that don’t generate income (or potential income).
- Based on the math of inflation, one has to earn 10% more every year to maintain the same standard of living (Assuming that the rate of inflation is 10%).
  - This issue can be tackled either by increasing income or decreasing expenses.
- Periods of austerity are good for innovation.
Anil Harwani described how a computer works based on the Von Neumann architecture to later discuss the differences between CPU and GPU.
- A modern CPU can have multiple cores which are good at performing general-purpose tasks, but a CPU is not good at performing extremely parallel operations.
  - Some CPUs might have accelerators for performing parallel operations on an order of magnitude of maybe 100s of threads, but they are not as good as GPUs.
- Every processor has I/O buses (could be PCIe (X4) links) that talks with NVMe(Non-Volatile Memory Express, a storage protocol running on top of PCIe) and a GPU can have a PCIe (x16) link, which translates into 4 times more bandwidth than the storage.
  - When a GPU is triggered, the OS triggers DMA (Direct Memory Access) which is a hardware mechanism that allows bypassing CPU so that the GPU can directly talk with the memory.
    
    We used GPUz to determine our GPU specifications. Interesting fact - This tool was built by a single person.
We discussed the differences between PCIe versions and the link transfer speeds that come with them.
- Speed differences are huge if we compare PC GPUs and commercial GPUs like NVIDIA A100.
  - This is a whitepaper that talks about the architecture of the A100 GPU.
Anil Harwani talked about the function of the voltage regulators on the GPU.
- Those little devices can suck power that is equivalent to or even more than a small refrigerator or an AC.
  - Images are available in this whitepaper.
We zoomed into GPU cores where we talked about shader modules which can essentially do fast math operations.
- Memory Controllers and L2 Cache help in loading the code in GPU which is then executed through the shader modules.
- All the shader cores are clubbed into units called Warps.
  - Each warp can execute 32 threads.
    
    Each thread can be a different instance of the same program.
Anil Harwani shared that he designed Nvidia’s DGX 100 which uses NVLink to connect 1120 GPUs which act as a single system.
- This system has a very high bandwidth, memory and lots of tensor cores.
- These types of systems are used to train things like ChatGPT.
- He also shared Rack Units and how they are used to allocate systems in a server.
- We discussed what a DIMM is.
We learned why mainframes are still used for handling transactions.
- Reason being - Business logic hasn’t changed for a long time.
- It is a very stable and trustworthy system.
- Anil Harwani shared a list of software practices required by nasa:
Singapore does not follow its geographical time zone mostly because of financial reasons.
- It tries to match the time with Japan which has adverse effects on the circadian rhythm of the people.
Anil Harwani suggested a podcast by BBC World Service - 13 minutes to the Moon
- Download link
- Spotify link

Projects Showcased

Rishit Dagli showcased 3D Transforms - a library to easily work with 3D data and make 3D transformations.
- Demo
- GitHub Repository
Tushar Nankanishowcased jsonexp a JS script for linking Twitter attendees using the attendance list generated by the chrome extension.
- It uses an attendee list to generate a map of attendees for attendees.adoc.
- GitHub Repository

Attendees

Meet Screenshot

For all summaries, please visit catchup.ourtech.community/summary.

Want to improve the content of this page?

Want to get more involved? Learn how to contribute.

We rise by lifting others.

— Robert Ingersoll

Kindness always wins.

— Selena Gomez

Alone we can do so little; together we can do so much.

— Helen Keller