Projects

A modern, extensible command-line tool for benchmarking GPU performance on NVIDIA CUDA devices.

Provides accurate, reproducible measurements of memory bandwidth, compute throughput, tensor core performance, kernel launch overhead, and reduction performance.

Supports multiple data types (FP32, FP16, BF16, INT8, FP4), comprehensive statistics, and multiple output formats (console, JSON, CSV).

An experimental network on real internet under two ASNs: AS151192 & AS199684, with maximum capacity of 5 Gbps.

It utilized self-developed VPN protocols and software to connect hundreds of nodes around the world, with profiling and telemetry.

Bird was used to set up BGP connections to upstreams and peers. It announced several IPv6 prefixes and hosted services for my use.