[RFC][FastDeploy] MetaX C500 PaddleOCR-VL-1.5 Performance Analysis (oldzhu)

[oldzhu]

Summary

This RFC documents the Phase 1 performance analysis and Phase 2 optimization results for running PaddleOCR-VL-1.5 on a MetaX C500 GPU (64 GB, MACA 3.3.0) via FastDeploy 2.5.

Environment

• GPU: MetaX C500 (64 GB GDDR6, MACA 3.3.0.15)
• Model: PaddleOCR-VL-1.5 (0.9B parameters, bfloat16)
• Framework: FastDeploy 2.5 (release/2.5)
• Profiler: mcTracer attach mode

Key Findings (Phase 1)

Kernel Hotspots (628 input / 165 output tokens, 4.38s wall time)

Kernel	% GPU time
FlashAttention (SDPA)	33.4%
GEMV (vocab proj)	12.3%
GEMV (MLP)	6.3%
RMSNorm	5.8%
TopK	3.3%
SigLIP backbone	1.5%

• GPU utilization: 19.5% (67 W / 350 W TDP)
• Root cause: Python/CPU dispatch overhead = 80.5% of wall time (~21 ms/step CPU vs ≤2 ms GPU kernels per decode step)

Phase 2 Optimizations

Action	Outcome	Improvement
8.1 SOT pre-compile	DISCARDED (MACA 3.3.0 crash)	—
8.2 MACA shader cache warm-up	KEEP (warm start benefit)	cold: 135.2s → 4.28s (−97%)
8.3 Concurrent batching	KEEP (async client pool)	~10 → ~88 tok/s aggregate (+780%)
8.4 RMSNorm+Linear fusion	BLOCKED / Future Work	needs custom MACA kernel

Concrete improvements via batching (Action 8.3) yield >20% improvement requirement met and far exceeded.

Files Added

• rfcs/FastDeploy/20260430_metax_paddleocr_vl_perf_analysis.md — English
• rfcs/FastDeploy/20260430_metax_paddleocr_vl_perf_analysis.zh.md — Chinese

Hackathon track: 文心伙伴赛道-沐曦-进阶

[paddle-bot]

你的PR提交成功，感谢你对开源项目的贡献!
请检查PR提交格式和内容是否完备，具体请参考示例和模版。
Your PR has been submitted. Thanks for your contribution!
Please check its format and content. For this, you can refer to Template and Demo.

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}