Why Are GPUs (Not) Fast - A Trip Through the Driver Stack
summary
Explains why GPUs trade latency for throughput: GPUs hide DRAM latency by massive multithreading and large register files, but register pressure reduces threads in flight and hurts performance; gives numbers (DRAM ~4,000 MT/s on a 32‑bit lane, ~20 ns access, bandwidth‑delay product 320 bytes) and shows how kernel vs user drivers (Mesa, EGL/OpenGL/Vulkan) batch commands to amortize syscall cost. Covers compositor pipelining, 60 Hz ≈16.6 ms frame latency, pipeline bubbles from shared data, and implicit vs explicit fences for dependency tracking.
AI-written summary · about
chapters (10)
- 0:00Intro
- 0:38Magic?
- 4:05Deep down (the memory lane)
- 6:33Throughput over latency
- 10:24GPU hardware
- 11:44GPU drivers
- 19:50Display composition
- 22:53Display pipelining
- 25:59Display latency reduction (failed)
- 33:05Bonus: fences
related talks
50mFrom Magic to Mechanics: Demystifying GPU Programming
30mUnderstanding GPUs
37mHigh Performance Computing: The Modern Approach
20mPassthrough/Headless GPU Gets Ahead
21mHPC Green Computing: Relationships Between Throughput and Latency
15mNixie: Efficient, Transparent Temporal Multiplexing for Consumer GPUs
23m1.2 Tb/S Packet Processing of Radio Telescope Data on a GPU
42mMaximizing Your Game's Performance in Unreal Engine