SM Architecture Overview: Inputs & Outputs

Inputs and outputs contract

Inputs & Outputs for SM Architecture Overview centers on SM IPC, functional-unit utilization, and front-end bubble ratio. The objective is to connect profiler evidence to root-cause mechanism and release-safe action.

Use this as the handoff contract across architecture, kernel, compiler, and silicon teams. Ambiguity here creates expensive late-stage rework.

INPUTS
  - workload definition and expected KPI target
  - kernel launch geometry, compiler flags, and toolchain versions
  - hardware assumptions: SM count, memory type, clocks, thermal envelope
  - acceptance criteria for throughput, latency tail, and stability

OUTPUTS
  - counter and timeline report with reproducible tags
  - bottleneck classification (compute, memory, fabric, thermal)
  - owner-signed mitigation proposal
  - benchmark rerun summary and release recommendation

Hierarchy map

GPU MEMORY HIERARCHY — SM Architecture Overview

                [ Registers ]
              latency:   1-2 cycles
                     |
                [ Shared/L1 ]
              latency:  20-40 cycles
                     |
                    [ L2 ]
              latency: 150-250 cycles
                     |
             [ HBM/GDDR VRAM ]
              latency: 300ns+ effective

Optimization lens: capacity vs latency

Scheduler map

WARP SCHEDULER VIEW — SM Architecture Overview

cycle ->      0    1    2    3    4
eligible   [W1,W2,W5] [W2] [W2,W7] [W7] [W3,W7]
issued         W1      W2    W7      W7    W3
stall reason    -    dep wait  -   mem wait  -

Scheduler objective: keep issue slots non-empty.
Focus: eligible warp quality

GPU deep dive

Shader-core throughput is gated by issue policy, register-bank access, and pipeline hazard behavior.

Concept diagram

SM CORE LOOP

warp schedulers -> issue ports -> ALU/FPU/Tensor pipelines
scoreboard + register file gate progress

Metric graph

SM BOTTLENECK MIX

dependency stalls   ███████
bank conflicts      ████
pipeline bubbles    ███

Reports and artifacts

• SM IPC dashboard

• issue stall taxonomy

• register-bank conflict log

• shader unit utilization

Mini case study

Compiler register allocation shifted operand banking, doubling RF conflicts and causing a 14% shader regression.

Debug branches

• Inspect scoreboard wait-depth trends

• Track RF conflicts by instruction class

• Separate front-end issue loss from backend saturation

Senior review question

Ask: which metric and benchmark pairing proves this topic is truly closed in production context?

Key takeaways

• Always pair micro-kernel metrics with end-to-end workload impact.

• Lock toolchain, driver, and launch metadata before comparing performance results.

Common pitfalls

• Optimizing occupancy without checking memory-system saturation.

• Comparing profiler captures from different driver or compiler builds.

• Declaring wins without reproducible accuracy and performance gates.

Handoff explanation

Inputs are not only API parameters or RTL configuration bits. For GPU design, inputs include workload distribution, launch geometry, shader/compiler form, memory layout, clocks, thermal state, SKU target, and runtime policy. Missing any of these makes the same counter mean different things.

Outputs must be decision-ready: SM IPC, functional-unit utilization, and front-end bubble ratio, the artifact set (SM block diagram, utilization heatmap, and issue-stage pipeline trace), a bottleneck class, owner, expected effect, and regression scope. A handoff that says only "performance improved" is not enough for architecture or silicon signoff.

The safest handoff format is a before/after packet: workload, revisions, counters, traces, root-cause hypothesis, chosen change, rejected alternatives, and rollback criteria.