Systems Papers
Systems paper note는 infrastructure, training efficiency, inference serving, distributed execution, agent tooling, reproducibility를 다룹니다.
Related Architecture Primitives
Some systems-heavy papers are kept in Architecture Papers when the contribution changes a core model primitive. Use this section as a bridge rather than duplicating the note.
| Paper | Why it matters for systems |
|---|---|
| FlashAttention | exact attention with lower HBM traffic and activation memory |
| GShard | sparse MoE architecture whose claims depend on automatic sharding, compiler partitioning, and distributed execution |
Reading Axes
- What bottleneck is addressed: memory, compute, communication, I/O, latency, or reliability?
- What is the system boundary: kernel, runtime, framework, scheduler, serving stack, or workflow?
- What workload is evaluated?
- Are baselines and hardware assumptions stated clearly?
- Can the result be generalized beyond the reported setup?
- Is the claim about latency, throughput, memory, cost, reliability, or developer workflow?
Concepts
- AI systems
- Training run
- Inference
- Model serving
- Latency and throughput
- Memory-compute tradeoff
- Reproducibility
- Retrieval-augmented generation
- Embedding retrieval
- Optimizer
- Learning rate schedule
- Batch size
- GPU memory
- Data loading and IO
- Distributed training
- Inference serving
- Reproducible run record
- Agent architecture
- Agent orchestration
- Agent evaluation