Reverse-Engineering the RTX 4090: A Deep Dive into GPU Memory Reads
CatAstro_Piyush · x · 2026-08-14
This in-depth technical post reverse-engineers the complete execution path of a global memory load instruction (LDG) through GPU hardware, specifically using an RTX 4090.
- Instruction Tracing: Starting with a simple vector-add kernel, it traces how the compiled SASS instruction computes an address in a register and requests data from memory.
- Hardware Hierarchy: It breaks down how a 4-byte request across 32 thread lanes traverses cache lines, address translations, crossbars, and ultimately reaches L2 cache slices and DRAM chips.
- Reverse Engineering: Because NVIDIA poorly documents many of these low-level details, the author runs timing experiments directly on the hardware to infer and verify the undocumented mechanics.
More from Infra
- Dev publicly asks neoclouds for discounts on a $30k GPU compute purchase — Sauers_ · 2026-10-16
- New hardware day: local LLM deployment the moment it arrives — lxfater · 2026-10-16
- Together AI's score centering fixes RL training collapse under quantized inference — nrehiew_ · 2026-10-16
- Inference is splitting prefill and decode across specialized chips, not just H100s — AccBalanced · 2026-10-16
- Rebasely launches in beta: Git hosting built for humans and coding agents — handotdev · 2026-10-16
- Drama3 TTS lands in audio.cpp: up to 5x real-time on RTX 5090, <8GB VRAM mode too — Acceptable-Cycle4645 · 2026-10-16