From 8c197439d74e15f5cac4b51fa9d7084d1d5a14fc Mon Sep 17 00:00:00 2001
From: Matthias Reso <13337103+mreso@users.noreply.github.com>
Date: Wed, 23 Apr 2025 16:06:25 -0700
Subject: [PATCH] Add benchmark values for pct_achievable_gemm_tops and
 pct_achievable_mem_bw for MI300X

---
 torchao/testing/float8/roofline_utils.py | 10 ++++------
 1 file changed, 4 insertions(+), 6 deletions(-)

diff --git a/torchao/testing/float8/roofline_utils.py b/torchao/testing/float8/roofline_utils.py
index 92becb9b94..f0b873c352 100644
--- a/torchao/testing/float8/roofline_utils.py
+++ b/torchao/testing/float8/roofline_utils.py
@@ -47,12 +47,10 @@
         "fp8_peak_tops": 2614e12,
         # 5.3 TB per second
         "peak_mem_bw_bytes_sec": 5.3e12,
-        # for now, copy over from H100
-        # TODO(future): run measurement on hardware
-        "pct_achievable_gemm_tops": 0.78,
-        # for now, copy over from H100
-        # TODO(future): run measurement on hardware
-        "pct_achievable_mem_bw": 0.92,
+        # based on microbenchmark (fw + bw gemms) with M,K,N = 3 * (8192,)
+        "pct_achievable_gemm_tops": 0.47,
+        # based on microbenchmark with pointwise triton kernel with large inputs
+        "pct_achievable_mem_bw": 0.72,
     },
     # TODO(future): more GPU names
 }