DANDELION

From FORTRAN to GPU Computing

I realized that the FORTRAN based simulation was performing quite slowly, which led me to research possible optimization techniques. In doing so, I discovered methods for parallelizing code on Intel processors and conducted several roofline model analyses to estimate potential performance gains. However, after careful consideration, I ultimately abandoned this approach. It was at that point that I became aware of GPU computing as an alternative. Consequently, all credit for my progress with CUDA belongs to the graphics card shown on the left.

SHARED MEMORY

After some research, I found that instead of relying on global memory for every operation, the data can be loaded serially into the L2 cache, the calculations can be performed there, and the results can then be written back to global memory before loading the next data set. This approach cuts the overall simulation time by approximately half

ARM Optimization Stack

• ARM CPU vectorization (Neon, SVE/SVE2)
• Cloud-optimized ARM processors (Graviton, Ampere, Grace)
• ARM HPC toolchain (compilers, performance libraries, debuggers)
• ARM GPU compute (Mali, OpenCL, Vulkan)
• Memory architecture and concurrency on ARM
• Performance analysis and profiling tools

• Automated Phase Classification: Train classifiers (Random Forest, CNN) on labeled data
• Physics Surrogate Models: O(N²) force calculation is the main bottleneck (>80% of compute time), Graph Neural Networks (GNN) learn particle interactions
• Intelligent Parameter Optimization: Bayesian optimization (Optuna) guides parameter selection
• Trajectory Prediction: Sequence models (LSTM, Transformer) learn dynamics

NEIGHBOUR LIST

After following a professor's lecture in which he explained the neighbor list algorithm used in molecular dynamics simulations, I implemented the same approach in my own work. The results demonstrated a substantial improvement in computational performance while preserving the accuracy of the simulation

Arrays, hashing, trees

Graph search toolkit

A dedicated section on ML was established

Multi-Agent Systems

• Python (coordination logic, task decomposition)
• Message queues (Redis, Kafka for inter-agent communication)
• Vector databases (HNSW for semantic memory search)
• SQLite/AgentDB (persistent memory storage)
• Reinforcement learning algorithms (pattern propagation)
• Circuit breakers and retry mechanisms (fault tolerance)

The Beginning

I was wandering through an electronics market, moving from shop to shop and asking vendors whether they had a component that could help me design a small electronic application. Deep inside the market, in a tiny and easily overlooked shop, one seller casually suggested, "Why don't you try programming this tiny chip?" I decided to give it a try. I built a few simple electronics circuits, sold them to my friends, earned a small amount of money and with that, I bought my first smartphone.

Edge Intelligence

Enables inference on ultra-low-power devices with strict power, memory, and cost constraints.