Run example 7 (gemm_all_scatter) and report benchmark results

[Copilot]

Executed examples/07_gemm_all_scatter/benchmark.py with default parameters across 8 AMD GPUs (MI300X class) in both validation and benchmark modes.

Results

• Matrix shape: M=8192, N=4608, K=36864, fp16
• Validation: ✅ passed on all 8 ranks
• Throughput: ~1706 TFLOPS (avg across ranks)
• Total kernel time: ~1.634 ms | GEMM kernel: ~1.443 ms (126 iterations)
• Block config: BLK_M=256, BLK_N=64, BLK_K=64 | 288 tiles/rank

[Iris] [0/8] tile matmul + all_scatter (total_tiles=288): 1.634 ms  1703.445 tflops
[Iris] [1/8] tile matmul + all_scatter (total_tiles=288): 1.629 ms  1708.400 tflops
[Iris] [2/8] tile matmul + all_scatter (total_tiles=288): 1.631 ms  1706.170 tflops
[Iris] [3/8] tile matmul + all_scatter (total_tiles=288): 1.630 ms  1707.457 tflops
[Iris] [4/8] tile matmul + all_scatter (total_tiles=288): 1.630 ms  1707.831 tflops
[Iris] [5/8] tile matmul + all_scatter (total_tiles=288): 1.630 ms  1707.091 tflops
[Iris] [6/8] tile matmul + all_scatter (total_tiles=288): 1.635 ms  1701.952 tflops
[Iris] [7/8] tile matmul + all_scatter (total_tiles=288): 1.630 ms  1707.155 tflops

---

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