Engineering / core
GPU execution and synchronization
Understand thread hierarchy, execution ordering, synchronization scope, and launch overhead before optimizing device programs.
Your route through engineering
Follow the path or begin with a question. Every lesson is open.
Your place, at your pace. Progress is optional and saved only in this browser when available.
Read / explore / reason
Inside this module
Module 01 core
GPU execution and synchronization
Understand thread hierarchy, execution ordering, synchronization scope, and launch overhead before optimizing device programs.
- 01
Threads, blocks, and warps: map work without assuming an order
Trace CUDA indices, partial warps, block independence, and the difference between launched threads and useful elements.
24 min - 02
SIMT divergence: count active lanes without assuming lockstep
Analyze branch paths within a warp and use explicit synchronization under Volta-style independent thread scheduling.
24 min - 03
SM residency: account for every binding resource
Calculate an explicit hypothetical residency bound using effective allocated registers, shared memory, threads, and block limits.
24 min - 04
Occupancy: a latency-hiding tool, not the performance score
Separate resident warp fractions from ready work, memory traffic, arithmetic intensity, and measured application throughput.
24 min - 05
Synchronization scopes: order the right memory for the right participants
Separate warp and block barriers, memory fences, atomic publication, and host-visible completion.
24 min - 06
Streams: order work without assuming overlap
Trace same-stream dependencies, independent streams, host buffer lifetime, and the configured default-stream mode.
24 min - 07
Events: connect streams without blocking the entire device
Trace record and wait placement, event timing boundaries, and safe buffer reuse.
24 min - 08
Launch overhead and graphs: amortize a repeated dependency structure
Compare declared submission budgets, graph setup costs, and device critical paths without promising universal acceleration.
24 min - 09
Cooperative grids: synchronize only within a supported launch
Calculate a hypothetical all-resident grid limit and explain why ordinary inter-block spinning can deadlock.
24 min - 10
Asynchronous errors: find the failing work, not just the reporting call
Separate API and launch checks from execution completion, then use scoped synchronization and sanitizer evidence to narrow a device fault.
24 min