GPU Design · All levels
Rasterization & Early-Z
Graphics Pipeline Architecture: Rasterization maps primitives to fragments while early depth/stencil tests cull occluded work before expensive shader execution.
What this topic teaches
Rasterization & Early-Z converts GPU architecture concepts into review-ready engineering decisions. Rasterization maps primitives to fragments while early depth/stencil tests cull occluded work before expensive shader execution. The practical goal is to tie counters and traces to a specific mechanism, owner, and closure action.
The senior-engineer question
When raster throughput, early-Z kill rate, and overdraw reduction shifts, can you prove whether the root cause is SIMT control flow, SM scheduling, memory traffic, interconnect pressure, or graphics stage imbalance?
SIMT EXECUTION — Rasterization & Early-Z
warp 0 lanes: 0 1 2 3 4 5 6 7 ... 31
active mask : 1 1 1 1 0 0 1 1 ... 1
instruction : IF branch taken on active lanes
cycle 10: issue warp 0
cycle 11: issue warp 3
cycle 12: warp 0 reconverges
Focus: anchor discussion in lane masks and warp progress
Metric tracked: raster throughput, early-Z kill rate, and overdraw reductionPicture the architecture
Begin with an architecture sketch before touching tuning knobs. These diagrams are for design reviews, interview whiteboards, and closure discussions.
Raster and early depth culling
RASTER + EARLY-Z FLOW
triangles -> setup -> raster tiles -> early-Z/depth-stencil -> fragment queue
| pass | fail |
| shade| cull |
Higher early-Z kill rate reduces fragment ALU pressure and memory traffic.Overdraw bottleneck lens
BANDWIDTH ROOFLINE — Rasterization & Early-Z
performance
^
| compute ceiling
| /
| /
|-------------/------------------ memory ceiling
+------------------------------------------> operational intensity
memory-bound compute-bound
Interpretation: show overdraw-driven memory pressure vs culling gainsSM and datapath context
SM BLOCK DIAGRAM — Rasterization & Early-Z
+---------------------------+
| Warp Schedulers / Dispatch|
+------------+--------------+
|
+---------------+----------------+
| Register File / Operand Cross |
+--------+---------------+-------+
| |
[ALU/FPU] [LD/ST]
| |
+-------+-------+
|
L1 / Shared Mem
Focus: front-end to execute dataflowMemory hierarchy context
GPU MEMORY HIERARCHY — Rasterization & Early-Z
[ 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 latencyScheduler context
WARP SCHEDULER VIEW — Rasterization & Early-Z
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 qualityOwnership layers
GPU OWNERSHIP LAYERS — Rasterization & Early-Z
artifact area owner
---------------- ----------------------------
architecture raster backend owner
RTL/microarch graphics architect
software/tools performance engineer
Rule: each metric needs a named owner before signoff.Evidence to collect
Primary metric: raster throughput, early-Z kill rate, and overdraw reduction.
Primary artifact: raster tile occupancy map, depth-test effectiveness report, and overdraw heatmap.
Owners to include: raster backend owner, graphics architect, performance engineer.
One reproducible failing workload and one stable comparator workload.
One counter capture that separates compute issue from memory/interconnect pressure.
Roofline lens
BANDWIDTH ROOFLINE — Rasterization & Early-Z
performance
^
| compute ceiling
| /
| /
|-------------/------------------ memory ceiling
+------------------------------------------> operational intensity
memory-bound compute-bound
Interpretation: identify compute vs memory boundCoalescing lens
COALESCING PATTERN — Rasterization & Early-Z
WARP ADDRESSES
lane: 0 1 2 3 4 5 6 7
addr: 0 4 8 C 10 14 18 1C -> contiguous -> 1 transaction segment
lane: 0 1 2 3 4 5 6 7
addr: 0 40 8 48 10 50 18 58 -> strided/scatter -> many segments
Effect: fewer coalesced segments => better bandwidth efficiency.
Focus: transaction inflation from scatterSubpages in this topic
Each topic includes mechanism, inputs/outputs, reports, debug, worked example, pitfalls, interview, checklist, theory deep dive, design space, case study, walkthrough, matrix, software view, and silicon impact.
Key takeaways
Always connect warp behavior to measured counters before proposing fixes.
Treat memory transaction quality as equal priority to compute utilization.
Close decisions with explicit owners and reproducible benchmark evidence.
Common pitfalls
Copying tuning patterns from unrelated workloads or scenes.
Using occupancy as a success metric without stall classification.
Declaring closure without end-to-end frame or kernel validation.
GPU deep dive
Frame-time stability depends on balancing fixed-function stages with programmable shader pressure.
Concept diagram
GRAPHICS PIPELINE
vertex -> tessellation -> raster -> fragment -> ROP/blendMetric graph
FRAME-TIME PRESSURE
fragment shading load ████████
raster backpressure █████
ROP/blend stalls ████Reports and artifacts
stage occupancy timeline
early-Z efficiency report
ROP queue depth
overdraw heatmap
Mini case study
Async compute overlapped with heavy fragment scenes and triggered ROP queue buildup, causing p99 frame spikes.
Debug branches
Correlate frame spikes with stage-level queues
Validate early-Z effectiveness under real content
Isolate graphics-compute arbitration conflicts
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.