← Curriculum

C02 Stage B · Core AI 90 h (40 T / 50 P) Statutory Carries 25% of QP weight

Model Evaluation & Performance Engineering

Measures, profiles and optimises a model to fit real system constraints, and documents the trade-off.

Prerequisite
C01
Modules
5
NOS
SSC/N8121
Statutory mapping
Model Curriculum Module 5 — delivers SSC/N8121 in full

Tech stack

cProfileline_profilermemory_profilerPyTorch profilerONNX

Modules

5 modules · 90 hours

M1 · Define model parameters

18 h · PC1, PC2
TOPICS
  • Technical specifications and limitations of the host system
  • CPU/GPU architecture, cores, clock, cache hierarchy
  • RAM and VRAM ceilings
  • Storage I/O
  • OS constraints
  • Data flows and structures: tensors, batches, feature stores
  • Serialisation formats
  • Precision (fp32/fp16/int8)
  • Interface contracts and schema definition

Lab: Author a formal specification sheet documenting host limits and the model's data-flow graph

Course material for this module is in production.

M2 · Runtime, speed and memory analysis

20 h · PC3
TOPICS
  • Algorithmic complexity and Big-O notation
  • Time and space trade-offs
  • Profiling tools
  • Latency percentiles (p50/p95/p99)
  • Throughput and cold start
  • Memory leaks, fragmentation, garbage collection
  • Batch-size versus memory curves

Lab: Profile a model end to end; produce a flame graph and a bottleneck report ranked by impact

Course material for this module is in production.

M3 · Parallel programming constraints

18 h · PC4
TOPICS
  • Flynn's taxonomy: SISD, SIMD, MISD, MIMD
  • Threads versus processes
  • The Python GIL
  • Multiprocessing and vectorisation
  • BLAS threading
  • GPU execution model, kernels, occupancy, memory bandwidth
  • Data versus model parallelism
  • Amdahl's law
  • Synchronisation, race conditions, deadlock

Lab: Parallelise an inference workload; measure scaling efficiency against Amdahl's prediction

Course material for this module is in production.

M4 · Optimise to fit constraints

20 h · PC5
TOPICS
  • Quantisation (post-training and quantisation-aware)
  • Pruning
  • Knowledge distillation
  • Operator fusion and graph optimisation
  • ONNX export
  • Batching strategies
  • Caching
  • Mixed precision
  • Accuracy-versus-resource trade-off analysis
  • Regression testing after optimisation

Lab: Reduce model latency and memory to a stated budget while holding accuracy within tolerance; present the trade-off curve

Course material for this module is in production.

M5 · Document and feed back

14 h · PC6, PC7
TOPICS
  • Model cards
  • Evaluation reports
  • Versioned documentation
  • Reproducibility metadata
  • Anomaly identification and referral
  • Structured performance feedback to the design team
  • Organisational documentation and data-sharing policy
  • Escalation paths

Lab: Publish a model card and deliver a formal performance-feedback review to the modelling team

Course material for this module is in production.

Real-world work scenario

Production inference must run within 200 ms at p95 on a 4-core CPU instance with 8 GB RAM, but the current model averages 850 ms. Establish the baseline, profile to locate the bottleneck, apply quantisation and batching, verify accuracy remains within the agreed tolerance, document the change in the model card, and formally report the trade-off to the data-science team.

Assessment

Theory 30 marks · Practical 70 marks (mirrors the statutory split). Practical is graded on measured before/after evidence, not narrative.