Example 07 – GEMM all-scatter: IntelliKit Metrix profiling results

[Copilot]

Runs the GEMM all-scatter benchmark (Example 07) on 2× AMD MI300X (gfx942) and profiles persistent_gemm_all_scatter with IntelliKit Metrix across memory and compute profiles at two problem sizes.

Results

Problem	Performance	Total Latency	Comm Overhead
M=4096, N=4096, K=4096	205.8 TFLOPS	0.668 ms	~21%
M=8192, N=4608, K=36864 (default)	612.2 TFLOPS	4.546 ms	~3%

Profiling Highlights (large problem, Metrix)

• Compute-bound — HBM arithmetic intensity ~274 FLOP/byte; HBM bandwidth utilization only ~3.4%
• L2 hit rate ~82% — effective data reuse across the K-reduction loop
• Coalescing efficiency 25%, global store efficiency 6.25% — scatter writes to remote ranks are the primary optimization target; increasing BLK_N would raise store granularity
• Zero LDS bank conflicts, zero atomic latency — shared memory and synchronization design is clean

Changes

• examples/07_gemm_all_scatter/profiling_results.md — full metric tables, benchmark numbers, and analysis
• .gitignore — exclude .rocprofv3/ profiling artifacts

---

💬 We'd love your input! Share your thoughts on Copilot coding agent in our 2 minute survey.