Skip to content

update roofline chart for qk_norm #123

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
Laurawly merged 3 commits into from
Mar 18, 2026
Merged

update roofline chart for qk_norm #123

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
72db4cb
update roofline chart
Laurawly Mar 18, 2026
3199f5b
asdf
Laurawly Mar 18, 2026
b29b061
asdf
Laurawly Mar 18, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
52 changes: 21 additions & 31 deletions oink/README.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -119,42 +119,32 @@ baseline here is a **Q/K-norm adaptation** derived from the
[CUTLASS CuTeDSL Blackwell RMSNorm example](https://github.com/NVIDIA/cutlass/blob/main/examples/python/CuTeDSL/blackwell/rmsnorm.py),
not the example kernel used unchanged.

Representative steady-state medians from one GB300 run are shown below
(absolute microseconds may vary slightly run to run, but the ranking and
trend were stable).

#### Q path (`N=8192`, `scale=3.87`)

| M | CUTLASS (us) | Oink (us) | Speedup |
|---:|---:|---:|---:|
| 1 | 1.4 | 1.2 | 1.12x |
| 32 | 1.9 | 1.4 | 1.39x |
| 128 | 3.3 | 1.6 | 2.00x |
| 512 | 7.5 | 2.7 | 2.74x |
| 1024 | 12.6 | 4.0 | 3.12x |
| 4096 | 47.3 | 16.4 | 2.88x |
| 8192 | 93.7 | 38.0 | 2.47x |
| 16384 | 186.1 | 76.0 | 2.45x |
| 32768 | 371.5 | 152.7 | 2.43x |

#### K path (`N=1024`, `scale=1.0`)

| M | CUTLASS (us) | Oink (us) | Speedup |
|---:|---:|---:|---:|
| 1 | 1.3 | 1.2 | 1.06x |
| 32 | 1.6 | 1.3 | 1.21x |
| 128 | 1.6 | 1.3 | 1.21x |
| 512 | 2.3 | 1.4 | 1.58x |
| 1024 | 3.3 | 1.6 | 2.03x |
| 4096 | 7.6 | 2.5 | 3.03x |
| 8192 | 12.8 | 3.8 | 3.33x |
| 16384 | 23.1 | 6.5 | 3.56x |
| 32768 | 47.1 | 16.4 | 2.87x |
For roofline context, we also plot the same workload using a dedicated
useful-bandwidth harness: median CUDA-event timing plus a logical IO model of
one read + one write of the fused `[M, N]` tensor. This is the physically
meaningful view for comparing against the measured practical GB300 BF16 stream
roof, whereas the steady-state CUDA-graph replay medians below are better read
as a latency view.

<div align="center">
<img src="benchmarks/media/gb300_bf16_qk_norm_oink_vs_cutedsl_roofline.svg" alt="GB300 BF16: Q/K-norm roofline (Oink vs CuTeDSL)">
</div>

Representative steady-state CUDA-graph replay medians from one GB300 run are
shown below (absolute microseconds may vary slightly run to run, but the
ranking and trend were stable).

- Q path: Oink is roughly **2.4–3.1x faster** than the CuTeDSL baseline on
representative multi-row workloads.
- K path: Oink is roughly **2.0–3.6x faster** on the same sweep.

Takeaways from the GB300 Q/K-norm sweep:

- For the user-relevant multi-row workloads, Oink beats the CuTeDSL/CUTLASS
baseline by comfortably more than 20%.
- In the roofline view, Oink gets close to the practical GB300 BF16 streaming
ceiling on the large-row Q/K shapes, while the CuTeDSL baseline stays much
farther from the roof.
- The only cases below 20% are the tiny single-row latency-floor microcases:
Q `M=1` is ~12% faster and K `M=1` is ~6% faster.
- Correctness spot-check from the same harness:
Expand Down
Loading
Loading