You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Copy file name to clipboardExpand all lines: docs/en/architecture/index.md
+14Lines changed: 14 additions & 0 deletions
Display the source diff
Display the rich diff
Original file line number
Diff line number
Diff line change
@@ -18,6 +18,12 @@ This repository is not organized as “one fast kernel plus a benchmark screensh
18
18
19
19
That structure makes the project useful for learning, review, and interviews: readers can explain **why** a kernel exists before they talk about how fast it is.
20
20
21
+
## What this repository is trying to prove
22
+
23
+
- SGEMM optimization should read as a reasoning chain, not a bag of isolated tricks.
24
+
- Performance claims only count when they stay attached to correctness policy and benchmark scope.
25
+
- Tensor Core acceleration is only persuasive when constraints and fallback behavior are explicit.
26
+
21
27
## System map
22
28
23
29
| Layer | Responsibility | Where to go next |
@@ -83,6 +89,14 @@ This is not just process documentation. It affects how the architecture is narra
83
89
4. Read [Tensor Core Path](/en/architecture/tensor-core-path) before interpreting WMMA benchmark numbers.
84
90
5. Use the existing kernel deep dives when you want implementation detail instead of system rationale.
85
91
92
+
## Fast reviewer path
93
+
94
+
1. Read this page for the system claim.
95
+
2. Read [Kernel Ladder](/en/architecture/kernel-ladder) for the optimization order.
96
+
3. Read [Validation Overview](/en/validation/) before trusting any benchmark claim.
97
+
4. Read [Methodology](/en/methodology/) when you need the concise explanation path used in reviews or interviews.
98
+
5. Use [Resources Hub](/en/resources/) to trace external sources and comparison points.
Copy file name to clipboardExpand all lines: docs/en/methodology/index.md
+17Lines changed: 17 additions & 0 deletions
Display the source diff
Display the rich diff
Original file line number
Diff line number
Diff line change
@@ -37,6 +37,23 @@ flowchart TD
37
37
3.**Use local GPU evidence early.** Diagnosis without runtime evidence usually misclassifies the bottleneck.
38
38
4.**Hand every speedup to Validation.** A gain only counts after correctness, scope, and reproducibility are re-checked.
39
39
40
+
## Interview-ready framing
41
+
42
+
Use this structure when you need to explain the project concisely in a review or interview:
43
+
44
+
1.**Problem framing** — SGEMM is a useful proxy for memory hierarchy, parallel mapping, and mixed-precision trade-offs.
45
+
2.**Optimization ladder** — each kernel changes one bottleneck class at a time.
46
+
3.**Correctness and trust model** — cuBLAS is the oracle, tolerances differ between FP32 and WMMA, and unsupported WMMA shapes fall back explicitly.
47
+
4.**Measurement discipline** — end-to-end and compute-only numbers are separated, and hosted CI is kept separate from local GPU proof.
48
+
5.**Engineering maturity** — unified launcher interfaces, RAII/error handling, mirrored docs, and OpenSpec governance show the work is maintained deliberately.
49
+
50
+
### High-value deep-dive questions
51
+
52
+
-**Why not just use cuBLAS?** Production code should, but this repository exists to demonstrate bottleneck diagnosis and defensible optimization reasoning.
53
+
-**Why is the Tensor Core path still behind cuBLAS?** The WMMA path here is educational and intentionally exposes staging, alignment, and fallback costs instead of hiding them behind a production-tuned stack.
54
+
-**How do you know the performance claims are trustworthy?** The project checks correctness first, labels benchmark scope explicitly, and separates CI-safe checks from local GPU runtime evidence.
55
+
-**What would you improve next for production work?** Architecture-specific launch tuning, deeper WMMA overlap, stronger profiler evidence, and CUTLASS-style maintainable kernels.
0 commit comments