granted07@kolkata:~$ whoami

shhh, the better scripter is speaking

cd ~/research

● running

gpu scheduling for hosted neural-network inference

Research on reducing latency when neural networks are served on GPUs.

Requests are modelled as a queue with batching, a finite buffer and cache hits.

Service times are measured on a real GPU and fed back into the simulator.

M / G(a,b) / 1 / N
  arrivals   exp(λ)
  service    gamma(μ(b), cov 0 · .5 · 1.4)
  batch b    1 2 4 8 16 32 64   # googlenet, onnx runtime cuda
  out        latency · blocking · little's law

A GPU answers in groups. How long should it wait for a fuller ferry?

cd ~/research && cat gamma →
gpu-queue

~/DES_GPUQueueingModel .c

A simulator of requests queuing on a GPU that runs a neural network.

Batch times come from real GoogLeNet runs, then get replayed under different loads.

open →
chess-engine

~/chess-engine .py

A chess engine in Python, written without a chess library.

Handles castling and en passant.

open →
nn-in-c

~/NeuralNetworkInC .c

A small neural network library in C.

Matrices, layers and training are written by hand, with no outside libraries.

open →
p2pchat

~/p2pchat .cpp

Connects two computers directly, with no server in between.

It asks the router to open a port using UPnP.

open →
mc-console

~/mc-console .ts

A web page that shows a Minecraft server's log live and lets you send it commands.

Each command passes a login token, a rate limit and input cleaning before it reaches the server.

open →
genuprising

~/genu-website .ts

The website for GenUprising, an advocacy group.

Built with Next.js and a Supabase database. Caching keeps pages under 200ms.

open →
repak-rebnk

~/repak-rebnk .rs

A fork of repak-rivals, a Rust library and command line tool.

It reads and writes the .pak archives that Unreal Engine games use.

open →

git remote -v

github.com/granted07