GPU Computing · Cluster

GPU Performance

Roofline reasoning and bottleneck diagnosis.

  1. 01 The Roofline Bound and Arithmetic Intensity How the ratio of arithmetic to bytes moved places a kernel under either a compute ceiling or a bandwidth ceiling, and why the ridge point differs between devices. foundational · Jul 24, 2026
  2. 02 Achieved Bandwidth Versus the Quoted Peak Why a measured stream rate falls short of the number on the datasheet, how to count the bytes a kernel really moved, and which part of the gap is structural. foundational · Feb 4, 2025
  3. 03 Occupancy as a Latency-Hiding Budget How resident warps are rationed by registers, shared memory, and block granularity, and how to compute the number of concurrent requests a given occupancy can sustain. foundational · Mar 5, 2026
  4. 04 Wave Quantization and the Tail Effect Why a grid that fills the device one-and-a-bit times wastes the surplus, how to compute the number of waves, and which grid shapes avoid a long final wave. intermediate · Sep 26, 2023
  5. 05 Warp Stall Reasons and Cycle Attribution How a profiler turns idle issue slots into named stall categories, what each category actually means at the scheduler, and how to read the breakdown without over-interpreting it. intermediate · Jan 22, 2026
  6. 06 Memory-Level Parallelism and Latency Tolerance How many independent memory requests a kernel keeps in flight determines whether it can saturate bandwidth, and how unrolling and load reordering raise that count without adding warps. intermediate · Nov 20, 2023
  7. 07 Sector Counting and Why Cache Hit Rate Misleads How memory transactions are accounted in sectors, why a high hit rate can coexist with poor performance, and how to read the traffic counters that separate the two. intermediate · Apr 6, 2022
  8. 08 Issue-Limited Kernels and the Cost of Divergence When the instruction stream rather than the memory system sets the runtime, how warp divergence turns one instruction into several, and how to tell the two apart with counters. intermediate · May 13, 2024
  9. 09 Kernel Fusion and the Cost of a Launch Boundary What a separate kernel launch actually costs in time and traffic, when merging two kernels into one is a net win, and which shapes cannot be fused. advanced · Dec 22, 2025
  10. 10 Clock Derating and the Roofline That Moves Why the ceiling a kernel is measured against depends on power state and duration, and how sustained clocks change which side of the ridge a kernel occupies. advanced · Jan 2, 2026